def main(args=None):
# handle inputs #
parser = argparse.ArgumentParser()
parser.add_argument("-s",
"--scene",
metavar='FILE',
default="./model/scene.xml",
help="XML scene file defining the model")
parser.add_argument(
'output_file',
nargs='?',
default=None,
help="Filename to write scene.kml to. Default is scene.kml in same dir"
)
options = parser.parse_args(args)
scene_xml, scene_kml = options.scene, options.output_file
scene = boxm2_scene_adaptor(scene_xml, "cpp")
description = adaptor.describe_scene(scene.scene)
for key, value in description.iteritems():
print(" key = {}, value = {}".format(key, value))
if scene_kml is None:
scene_kml = os.path.join(description["dataPath"], "scene.kml")
adaptor.write_scene_to_kml(scene.scene, scene_kml)
scene.clear_cache()
def updateScene(self, stepNum):
# Should initialize a GPU
scenePath = os.path.abspath(self.config.uSceneFileName)
scene = boxm2_scene_adaptor.boxm2_scene_adaptor(scenePath, "gpu0")
# Get list of imgs and cams
imgsDir = os.path.abspath(
os.path.join(self.config.NVMOutputDirName, "imgs"))
camsDir = os.path.abspath(
os.path.join(self.config.NVMOutputDirName, "cams"))
if False == os.path.isdir(imgsDir):
log("The folder %s does not exist." % (imgsDir))
return -stepNum
if False == os.path.isdir(camsDir):
log("The folder %s does not exist." % (camsDir))
return -stepNum
imgFiles = getAllFilesInSubDirs(imgsDir)
camFiles = getAllFilesInSubDirs(camsDir)
if len(imgFiles) != len(camFiles):
log("The number of files in %s is %d." % (imgsDir, len(imgFiles)))
log("The number of fiels in %s is %d." % (camsDir, len(camFiles)))
log("These numbers should match and they do not.")
return -stepNum
if len(imgFiles) == 0:
log("No image files were found in %s." % (imgsDir))
return -stepNum
imgFiles.sort()
camFiles.sort()
# Make two passes over the image set
for p in xrange(0, self.config.updateScenePassCount):
frames = range(0, len(imgFiles), 1)
if self.config.randomizeUpdateOrder:
random.shuffle(frames)
for idx, i in enumerate(frames):
pcam = vpgl.load_perspective_camera(
os.path.join(camsDir, camFiles[i]))
img, ni, nj = vil.load_image(os.path.join(
imgsDir, imgFiles[i]))
scene.update(pcam, img, 1, "", "gpu0")
if 0 == (
idx % 15
): # No idea what is special about the magic number 15
scene.refine()
scene.write_cache()
return 0
def fetch_voxel_world_bounding_box(request, voxel_world_id): import boxm2_scene_adaptor import json import os voxel_world = voxel_globe.meta.models.VoxelWorld.objects.get(id=voxel_world_id) scene = boxm2_scene_adaptor.boxm2_scene_adaptor(os.path.join(voxel_world.directory, 'scene.xml'), 'cpp') return HttpResponse(json.dumps(scene.bbox))
def filter_number_observations(self, voxel_world_id, mean_multiplier=3.0,
history=None):
from boxm2_scene_adaptor import boxm2_scene_adaptor
import voxel_globe.meta.models as models
voxel_world = models.VoxelWorld.objects.get(id=voxel_world_id)
scene_file = os.path.join(voxel_world.directory, 'scene.xml')
scene = boxm2_scene_adaptor(scene_file, env['VIP_OPENCL_DEVICE'])
self.update_state(state='PROCESSING', meta={'stage':'remove low observers'})
scene.remove_low_nobs(mean_multiplier)
self.update_state(state='SAVING')
scene.write_cache()
def filter_number_observations(self,
voxel_world_id,
mean_multiplier=3.0,
history=None):
from boxm2_scene_adaptor import boxm2_scene_adaptor
import voxel_globe.meta.models as models
voxel_world = models.VoxelWorld.objects.get(id=voxel_world_id)
scene_file = os.path.join(voxel_world.directory, 'scene.xml')
scene = boxm2_scene_adaptor(scene_file, env['VIP_OPENCL_DEVICE'])
self.update_state(state='PROCESSING',
meta={'stage': 'remove low observers'})
scene.remove_low_nobs(mean_multiplier)
self.update_state(state='SAVING')
scene.write_cache()
def main():
import matplotlib
args = parse_args()
plot_scene = PlotScene()
if args.limits:
xyz = [float(i) for i in args.limits]
plot_scene.set_limits(xyz[0], xyz[1],\
xyz[2], xyz[3],\
xyz[4], xyz[5])
z_min = None
if args.scene:
scene = boxm2_scene_adaptor(args.scene, 'cpp')
plot_scene.draw_scene_box(scene)
z_min = scene.bbox[0][2]
if (args.limits and not z_min):
z_min = xyz[4]
if args.cameras:
cameras = []
camera_files = [
x for y in map(lambda x: glob(x), args.cameras) for x in y
]
camera_files = natural_sorted(camera_files)
for camera_file in camera_files:
krt = Krt.load(camera_file)
cameras.append(krt)
plot_scene.draw_cameras(cameras, z_min)
if args.cameras and args.diff:
cameras = []
camera_files = [
x for y in map(lambda x: glob(x), args.diff) for x in y
]
camera_files = natural_sorted(camera_files)
for camera_file in camera_files:
krt = Krt.load(camera_file)
cameras.append(krt)
plot_scene.draw_cameras(cameras, z_min, 'g')
plt.show()
def filter_number_observations(self, voxel_world_id, mean_multiplier=3.0):
from boxm2_scene_adaptor import boxm2_scene_adaptor
import voxel_globe.meta.models as models
import voxel_globe.tools
from vsi.tools.dir_util import copytree
voxel_world = models.VoxelWorld.objects.get(id=voxel_world_id)
with voxel_globe.tools.storage_dir('voxel_world') as storage_dir:
copytree(voxel_world.directory, storage_dir)
voxel_world.id = None
voxel_world.directory = storage_dir
voxel_world.name = 'Filtered at %d %s' % (mean_multiplier,
voxel_world.name)
voxel_world.service_id = self.request.id
voxel_world.save()
scene_file = os.path.join(storage_dir, 'scene.xml')
scene = boxm2_scene_adaptor(scene_file, env['VIP_OPENCL_DEVICE'])
self.update_state(state='PROCESSING', meta={'stage':'remove low observers'})
scene.remove_low_nobs(mean_multiplier)
self.update_state(state='SAVING')
scene.write_cache()
def filter_number_observations(self, voxel_world_id, mean_multiplier=3.0):
from boxm2_scene_adaptor import boxm2_scene_adaptor
import voxel_globe.meta.models as models
import voxel_globe.tools
from vsi.tools.dir_util import copytree
voxel_world = models.VoxelWorld.objects.get(id=voxel_world_id)
with voxel_globe.tools.storage_dir('voxel_world') as storage_dir:
copytree(voxel_world.directory, storage_dir)
voxel_world.id = None
voxel_world.directory = storage_dir
voxel_world.name = 'Filtered at %d %s' % (mean_multiplier,
voxel_world.name)
voxel_world.service_id = self.request.id
voxel_world.save()
scene_file = os.path.join(storage_dir, 'scene.xml')
scene = boxm2_scene_adaptor(scene_file, env['VIP_OPENCL_DEVICE'])
self.update_state(state='PROCESSING',
meta={'stage': 'remove low observers'})
scene.remove_low_nobs(mean_multiplier)
self.update_state(state='SAVING')
scene.write_cache()
def create_height_map(self, voxel_world_id, render_height):
import shutil
import urllib
import numpy as np
import brl_init
from boxm2_scene_adaptor import boxm2_scene_adaptor
from boxm2_adaptor import ortho_geo_cam_from_scene, scene_lvcs, scene_bbox
from vpgl_adaptor_boxm2_batch import convert_local_to_global_coordinates, geo2generic, save_geocam_to_tfw
from vil_adaptor_boxm2_batch import save_image, scale_and_offset_values, stretch_image, image_range
import vsi.io.image
import voxel_globe.tools
import voxel_globe.tools.hash
import voxel_globe.tools.camera
import voxel_globe.meta.models as models
import voxel_globe.ingest.payload.tools
with voxel_globe.tools.task_dir('height_map', cd=True) as processing_dir:
voxel_world = models.VoxelWorld.objects.get(id=voxel_world_id)
scene = boxm2_scene_adaptor(os.path.join(voxel_world.directory, 'scene.xml'), env['VIP_OPENCL_DEVICE'])
ortho_camera, cols, rows = ortho_geo_cam_from_scene(scene.scene)
tfw_camera = os.path.join(processing_dir, 'cam.tfw')
save_geocam_to_tfw(ortho_camera, tfw_camera)
with open(tfw_camera, 'r') as fid:
geo_transform = [float(x) for x in fid.readlines()]
(x0,y0,z0),(x1,y1,z1) = scene_bbox(scene.scene)
lvcs = scene_lvcs(scene.scene)
#lvcs = vpgl_adaptor.create_lvcs(lat=origin[1], lon=origin[0], el=origin[2],
# csname="wgs84")
_,_,min_height = convert_local_to_global_coordinates(lvcs, x0, y0, z0)
if render_height is None:
render_height = z1+(z1-z0)/1000
#z1+(z1-z0)/1000 is basically to say "just a little above the top" *2 is
#1) overkill and 2) doesn't work with sign, +1 could go crazy in an
#arbitrarily scaled system, so this just calculates ".1% more" which is
#more than good enough
else:
render_height = render_height - voxel_world.origin[2]
logger.critical("Render Height is %f (%s)", render_height, type(render_height))
generic_camera = geo2generic(ortho_camera, cols, rows, render_height, 0)
z_exp_img, z_var_img = scene.render_z_image(generic_camera, cols, rows)
#This is TECHNICALLY wrong, it assumes the earth is flat.
scale_and_offset_values(z_exp_img, 1, min_height)
height_filename = os.path.join(processing_dir, 'height.tif')
save_image(z_exp_img, height_filename)
checksum = voxel_globe.tools.hash.sha256_file(height_filename)
with voxel_globe.tools.image_sha_dir(checksum) as image_dir:
original_filename = os.path.join(image_dir, 'height_map.tif')
#If the exact file exist already, don't ingest it again. Unlikely
if not os.path.exists(original_filename):
img = vsi.io.image.imread(height_filename)
vsi.io.image.imwrite_geotiff(img.raster(), original_filename,
[geo_transform[x] for x in [4,0,1,5,2,3]])
zoomify_filename = os.path.join(image_dir, 'zoomify.tif')
img_min, img_max = image_range(z_exp_img)
if img_min == img_max:
zoomify_image = z_exp_img #At least it won't crash
else:
zoomify_image = stretch_image(z_exp_img, img_min, img_max, 'byte')
save_image(zoomify_image, zoomify_filename)
zoomify_name = os.path.join(image_dir, 'zoomify')
voxel_globe.ingest.payload.tools.zoomify_image(zoomify_filename, zoomify_name)
img = voxel_globe.meta.models.Image(
name="Height Map %s (%s)" % (voxel_world.name,
voxel_world.id),
image_width=cols, image_height=rows,
number_bands=1, pixel_format='f', file_format='zoom',
service_id=self.request.id)
img.filename_path=original_filename
img.save()
image_set = models.ImageSet.objects.get_or_create(name="Height Maps",
defaults={"_attributes":'{"autogen":true}'})[0]
image_set.images.add(img)
gsd = scene.description['voxelLength']
camera_center = ((x0+x1)/2, (y0+y1)/2, z1+10000)
d = z1-z0+10000
k=np.eye(3)
k[0,2] = cols/2
k[1,2] = rows/2
k[0,0] = k[1,1] = d/gsd
r=np.eye(3)
r[0,0]=-1
t = -r.T.dot(camera_center)
camera=voxel_globe.tools.camera.save_krt(self.request.id, img, k, r, t,
voxel_world.origin)
camera_set=voxel_globe.meta.models.CameraSet(\
name="Height Map %s (%s)" % (voxel_world.name, voxel_world.id), \
images=image_set, service_id=self.request.id)
camera_set.save()
camera_set.cameras.add(camera)
def run_build_voxel_model_bp(self, image_set_id, camera_set_id, scene_id, bbox,
skip_frames, cleanup=True):
import random
import glob
import math
import numpy as np
from vsi.tools.redirect import StdRedirect
from voxel_globe.meta import models
from voxel_globe.tools.camera import get_krt
import voxel_globe.tools
from boxm2_scene_adaptor import boxm2_scene_adaptor
import brl_init
from vil_adaptor_boxm2_batch import load_image
import vpgl_adaptor_boxm2_batch as vpgl
from vsi.vxl.create_scene_xml import create_scene_xml
from vsi.tools.dir_util import copytree
from vsi.tools.file_util import lncp
def rectint(recta,rectb):
lx = max(recta[0],rectb[0])
rx = min(recta[2],rectb[2])
by = max(recta[1],rectb[1])
ty = min(recta[3],rectb[3])
if lx > rx or by > ty :
return [0,0,0,0],0
else:
return [lx,by,rx,ty], (rx-lx)*(ty-by)
def generate_subsetim(scene,camfiles,ni,nj):
subsetIdx = []
refIndices = []
minDepOverlap = 0.25
minRefOverlap = 0.5
minIndepAngle = 5.0
minRefAngle = 5.0
maxRefAngle = 15.0
minRefIndepAngle = 5.0
cosMinIndepAngle = math.cos( minIndepAngle*math.pi/180.0 );
cosMinRefAngle = math.cos( minRefAngle*math.pi/180.0 );
cosMaxRefAngle = math.cos( maxRefAngle*math.pi/180.0 );
cosMinRefIndepAngle = math.cos( minRefIndepAngle*math.pi/180.0 );
bbox = scene.bbox
grect=[scene.bbox[0][0],scene.bbox[0][1],scene.bbox[1][0],scene.bbox[1][1]]
worldoverlaps = []
camrects = []
cams = []
princAxis = []
for camfile in camfiles:
pcam = vpgl.load_perspective_camera(camfile)
prx,pry,prz=vpgl.get_backprojected_ray(pcam,ni/2,nj/2)
princAxis.append([prx,pry,prz])
Hmat = vpgl.compute_camera_to_world_homography(pcam,[0,0,1,-bbox[0][2]])
H = np.array(Hmat).reshape([3,3])
ps = np.dot(H,np.transpose([[0,0,1],
[ni,0,1],
[ni,nj,1],
[0,nj,1]]))
xs = ps[0,:]/ps[2,:]
ys = ps[1,:]/ps[2,:]
rect = [min(xs),min(ys),max(xs),max(ys)]
area = (rect[2]-rect[0])*(rect[3]-rect[1])
crect,carea = rectint(rect,grect)
#print crect,carea
if ( carea > 0 ):
cams.append(pcam)
camrects.append(crect)
worldoverlaps.append(carea/area)
usedcams = [False]*len(cams)
for i in range(0,len(cams)):
randidx = random.randint(0,len(cams)-1)
while usedcams[randidx]:
randidx = (randidx+1)%len(cams)
usedcams[randidx]= True
dep = False
for c2 in range(0,len(subsetIdx)):
cosAngle = np.dot(princAxis[randidx], princAxis[subsetIdx[c2]] )
if cosAngle > cosMinIndepAngle :
rectc2 = camrects[subsetIdx[c2]]
overlap,oarea = rectint(camrects[randidx] , rectc2)
tarea = (rectc2[2]-rectc2[0])*(rectc2[3]-rectc2[1])
if( oarea/tarea > minDepOverlap ):
dep = True
break
if dep:
continue
theseRefIndices= []
for c3 in range(0,len(cams)):
#Check angle disparity
cosAngle2 = np.dot(princAxis[randidx],princAxis[c3] );
if( cosAngle2 > cosMinRefAngle or cosAngle2 < cosMaxRefAngle ):
continue
# Check that a similar viewpoint isn't already used for reference
refDep = False
for c4 in range(0,len(theseRefIndices)):
#Check angle disparity
cosAngle3 = np.dot(princAxis[theseRefIndices[c4]],princAxis[c3] );
if( cosAngle3 > cosMinRefIndepAngle ):
refDep = True
break
#If similar viewpoint don't add
if( refDep ):
continue
theseRefIndices.append(c3)
#If at least one reference image save this viewpoint
if len(theseRefIndices) > 0 :
subsetIdx.append( randidx );
refIndices.append( theseRefIndices );
return subsetIdx, refIndices
def update_bp(scene, images, cameras, do_update_image=True, do_update_hmap=False):
_, ni, nj = load_image (images[0])
frames,refimages = generate_subsetim(scene,cameras,ni,nj)
for file_name in glob.glob(os.path.join(scene.model_dir, 'boxm2_*.bin')):
os.remove(file_name)
scene.init_uniform_prob()
sradius = 16
idents = []
weights = []
if do_update_image:
idents.append("if")
weights.append(1.0)
if do_update_hmap:
idents.append("hf")
weights.append(2.0)
for idx, i in enumerate(frames):
if do_update_image:
print "Iteration ",idx, "Image " , images[i];
####load image and camera
viewid = os.path.splitext(os.path.basename(images[i]))[0]
#### forming an app model using the neighbor images
for lindex in refimages[idx]:
lcam = vpgl.load_perspective_camera(cameras[lindex]);
limg, ni, nj = load_image (images[lindex]);
scene.update(lcam, limg,False, True,None ,"gpu0",0.05,viewid)
scene.update_if(False,viewid) # subtracting the image factor
scene.fuse_factors(idents,weights)
pcam = vpgl.load_perspective_camera(cameras[i]);
img, ni, nj = load_image (images[i]);
scene.compute_pre_post(pcam, img,viewid,100000,100000); # computing the new image factor
scene.update_if(True,viewid) # adding the image factor
scene.fuse_factors(idents,weights)
if do_update_hmap and idx % 2 == 0:
scene.update_hf(False) # subtracting the height-map factor
scene.fuse_factors(idents,weights)
zimg,zvar,ximg,yimg,probimg = scene.render_height_map()
#save_image(zimg, "./zimg.tif")
scene.compute_hmapf(zimg,zvar,ximg,yimg,sradius) # computing the height-map factor
scene.update_hf(True) # adding the height-map factor
scene.fuse_factors(idents,weights)
scene.write_cache()
def refine(scene):
scene.refine(0.3)
for filename in glob.glob(os.path.join(scene.model_dir, '[a-b]*.bin')):
os.remove(filename)
scene.write_cache()
with StdRedirect(open(os.path.join(voxel_globe.tools.log_dir(),
self.request.id)+'_out.log', 'w'),
open(os.path.join(voxel_globe.tools.log_dir(),
self.request.id)+'_err.log', 'w')):
openclDevice = os.environ['VIP_OPENCL_DEVICE']
opencl_memory = os.environ.get('VIP_OPENCL_MEMORY', None)
if opencl_memory:
opencl_memory = int(opencl_memory)
scene = models.Scene.objects.get(id=scene_id)
imageSet = models.ImageSet.objects.get(id=image_set_id)
imageList = imageSet.images.all()
with voxel_globe.tools.task_dir('voxel_world') as processing_dir:
logger.warning(bbox)
create_scene_xml(openclDevice, 0, float(bbox['voxel_size']),
lvcs1=(float(bbox['x_min']), float(bbox['y_min']),
float(bbox['z_min'])),
lvcs2=(float(bbox['x_max']), float(bbox['y_max']),
float(bbox['z_max'])),
origin=scene.origin, model_dir='.', number_bins=1,
output_file=open(os.path.join(processing_dir, 'scene.xml'), 'w'),
n_bytes_gpu=opencl_memory)
counter = 1
imageNames = []
cameraNames = []
os.mkdir(os.path.join(processing_dir, 'local'))
#Prepping
self.update_state(state='INITIALIZE', meta={'image_set_name': imageSet.name,
'stage':'camera fetch'})
for image in imageList:
(K,R,T,o) = get_krt(image, camera_set_id)
krtName = os.path.join(processing_dir, 'local', 'frame_%05d.krt' % counter)
with open(krtName, 'w') as fid:
print >>fid, (("%0.18f "*3+"\n")*3) % (K[0,0], K[0,1], K[0,2],
K[1,0], K[1,1], K[1,2], K[2,0], K[2,1], K[2,2])
print >>fid, (("%0.18f "*3+"\n")*3) % (R[0,0], R[0,1], R[0,2],
R[1,0], R[1,1], R[1,2], R[2,0], R[2,1], R[2,2])
print >>fid, ("%0.18f "*3+"\n") % (T[0,0], T[1,0], T[2,0])
imageName = image.filename_path
extension = os.path.splitext(imageName)[1]
localName = os.path.join(processing_dir, 'local',
'frame_%05d%s' % (counter, extension))
lncp(imageName, localName)
counter += 1
imageNames.append(localName)
cameraNames.append(krtName)
variance = 0.06
vxl_scene = boxm2_scene_adaptor(os.path.join(processing_dir, "scene.xml"),
openclDevice)
# loaded_imgs = []
# loaded_cams = []
# for i in range(0, len(imageNames), skip_frames):
# logger.debug("i: %d img name: %s cam name: %s", i, imageNames[i],
# cameraNames[i])
# self.update_state(state='PRELOADING', meta={'image_set_name': imageSet.name,
# 'stage':'image load',
# 'i':i,
# 'total':len(imageNames)})
# img, ni, nj = load_image(imageNames[i])
# loaded_imgs.append(img)
# pcam = load_perspective_camera(cameraNames[i])
# loaded_cams.append(pcam)
refine_cnt = 2
for rfk in range(0, refine_cnt, 1):
self.update_state(state='PROCESSING', meta={'image_set_name': imageSet.name,
'stage':'update 1'})
update_bp(vxl_scene, imageNames, cameraNames)
# self.update_state(state='PROCESSING', meta={'image_set_name': imageSet.name,
# 'stage':'update 2'})
# update_bp(vxl_scene, imageNames, cameraNames, True, True)
# self.update_state(state='PROCESSING', meta={'image_set_name': imageSet.name,
# 'stage':'update 3'})
# update_bp(vxl_scene, imageNames, cameraNames, True, True)
if rfk < refine_cnt-1:
self.update_state(state='PROCESSING',
meta={'image_set_name': imageSet.name,
'stage':'refine', 'i':rfk+1,
'total':refine_cnt})
refine(vxl_scene)
#Update color appearance
with open(os.path.join(processing_dir, "scene_color.xml"), 'w') as fid:
lines = open(os.path.join(processing_dir, "scene.xml"),
'r').readlines()
lines = [line.replace('boxm2_mog3_grey',
'boxm2_gauss_rgb').replace(
'boxm2_num_obs',
'boxm2_num_obs_single') for line in lines]
fid.writelines(lines)
vxl_scene = boxm2_scene_adaptor(os.path.join(processing_dir,
"scene_color.xml"),
openclDevice)
for idx, (image_name, camera_name) in enumerate(zip(imageNames, cameraNames)):
self.update_state(state='PROCESSING', meta={
'image_set_name': imageSet.name,
'stage':'color_update',
'i':idx+1, 'total':len(imageNames),
'images':len(imageNames)})
img, _, _ = load_image(image_name)
pcam = vpgl.load_perspective_camera(camera_name)
logger.debug("color_paint idx: %d", idx)
vxl_scene.update(pcam,img,False,False,None,openclDevice,
tnear = 1000.0, tfar = 100000.0)
vxl_scene.write_cache()
with voxel_globe.tools.storage_dir('voxel_world') as voxel_world_dir:
copytree(processing_dir, voxel_world_dir, ignore=lambda x,y:['local'])
models.VoxelWorld(
name='%s world (%s)' % (imageSet.name, self.request.id),
origin=scene.origin,
directory=voxel_world_dir,
service_id=self.request.id).save()
return {"image_set_name" : imageSet.name}
def generate_error_point_cloud(self, voxel_world_id, prob=0.5, history=None):
from glob import glob
import json
import numpy as np
from boxm2_adaptor import load_cpp, render_depth, cast_3d_point, \
cast_3d_point_pass2, write_cache, \
create_stream_cache
from boxm2_scene_adaptor import boxm2_scene_adaptor
from vpgl_adaptor import create_perspective_camera_krt, persp2gen, \
compute_direction_covariance
from boxm2_mesh_adaptor import batch_compute_3d_points, gen_error_point_cloud
from vsi.tools.redirect import Redirect, Logger as LoggerWrapper
import voxel_globe.tools
import voxel_globe.meta.models as models
from voxel_globe.tools.camera import get_krt
with Redirect(stdout_c=LoggerWrapper(logger, lvl=logging.INFO),
stderr_c=LoggerWrapper(logger, lvl=logging.WARNING)):
self.update_state(state='SETUP', meta={'pid':os.getpid()})
voxel_world = models.VoxelWorld.objects.get(id=voxel_world_id)
service_inputs = json.loads(voxel_world.service.inputs)
image_collection = models.ImageCollection.objects.get(
id=service_inputs[0][0])
images = image_collection.images.all()
scene = models.Scene.objects.get(id=service_inputs[0][1])
voxel_world_dir = voxel_world.directory
scene_filename = os.path.join(voxel_world_dir, 'scene.xml')
opencl_device = os.environ['VIP_OPENCL_DEVICE']
scene_gpu = boxm2_scene_adaptor(scene_filename, opencl_device)
scene_cpp = boxm2_scene_adaptor(scene_filename, 'cpp')
type_id_fname = "type_names_list.txt"
image_id_fname = "image_list.txt"
std_dev_angle = 0.1
cov_c_path = 'cov_c.txt'
cov_c = 0*np.eye(3)*0.8**2
with voxel_globe.tools.task_dir('generate_error_point_cloud', cd=True) \
as processing_dir:
np.savetxt(cov_c_path, cov_c)
for index, image in enumerate(images):
self.update_state(state='PROCESSING',
meta={'stage':'casting', 'image':index+1,
'total':len(images)})
k,r,t,o = get_krt(image.history(history))
perspective_camera = create_perspective_camera_krt(k, r, t)
(depth_image, variance_image, _) = render_depth(scene_gpu.scene,
scene_gpu.opencl_cache, perspective_camera,
image.imageWidth, image.imageHeight,
scene_gpu.device)
cov_v_path = 'cov_%06d.txt' % index
appearance_model = 'image_%06d' % index
generic_camera = persp2gen(perspective_camera, image.imageWidth,
image.imageHeight)
compute_direction_covariance(perspective_camera, std_dev_angle,
cov_v_path)
cast_3d_point(scene_cpp.scene,scene_cpp.cpu_cache,perspective_camera,
generic_camera,depth_image,variance_image,appearance_model)
cast_3d_point_pass2(scene_cpp.scene,scene_cpp.cpu_cache,generic_camera,
appearance_model,cov_c_path,cov_v_path)
write_cache(scene_cpp.cpu_cache, 1)
self.update_state(state='PROCESSING',
meta={'stage':'compute error'})
with open(image_id_fname, 'w') as fid:
print >>fid, len(images)
for index, image in enumerate(images):
print >>fid, 'image_%06d' % (index)
with open(type_id_fname,"w") as fid:
print >>fid, 2
print >>fid, "boxm2_point"
print >>fid, "boxm2_covariance"
mem=3.0
stream_cache = create_stream_cache(scene_cpp.scene, type_id_fname,
image_id_fname, mem)
batch_compute_3d_points(scene_cpp.scene, scene_cpp.cpu_cache, stream_cache)
self.update_state(state='EXPORTING',
meta={'stage':'ply'})
with voxel_globe.tools.storage_dir('generate_error_point_cloud') \
as storage_dir:
ply_filename = os.path.join(storage_dir, 'model.ply')
gen_error_point_cloud(scene_cpp.scene, scene_cpp.cpu_cache,
ply_filename, 0.5, True)
potree_filename = os.path.join(storage_dir, 'potree.ply')
with voxel_globe.tools.image_dir('point_cloud') as potree_dir:
convert_ply_to_potree(ply_filename, potree_dir)
models.PointCloud.create(name='%s point cloud' % image_collection.name,
service_id=self.request.id, origin=voxel_world.origin,
potree_url='%s://%s:%s/%s/point_cloud/%s/cloud.js' % \
(env['VIP_IMAGE_SERVER_PROTOCOL'], env['VIP_IMAGE_SERVER_HOST'],
env['VIP_IMAGE_SERVER_PORT'], env['VIP_IMAGE_SERVER_URL_PATH'],
os.path.basename(potree_dir)),
directory=storage_dir).save()
voxel_files = lambda x: glob(os.path.join(voxel_world_dir, x))
cleanup_files = []
cleanup_files += voxel_files('boxm2_covariance_*.bin')
cleanup_files += voxel_files('boxm2_point_*.bin')
cleanup_files += voxel_files('float16_image_*.bin')
for cleanup_file in cleanup_files:
os.remove(cleanup_file)
def run_build_voxel_model_bp(self,
image_collection_id,
scene_id,
bbox,
skip_frames,
cleanup=True,
history=None):
from distutils.dir_util import remove_tree
from shutil import move
import random
from vsi.tools.redirect import StdRedirect, Logger as LoggerWrapper
from voxel_globe.meta import models
from voxel_globe.tools.camera import get_krt
import voxel_globe.tools
from boxm2_scene_adaptor import boxm2_scene_adaptor
from vil_adaptor import load_image
from vpgl_adaptor import load_perspective_camera
from voxel_globe.tools.wget import download as wget
from vsi.vxl.create_scene_xml import create_scene_xml
from vsi.tools.dir_util import copytree
with StdRedirect(
open(
os.path.join(voxel_globe.tools.log_dir(), self.request.id) +
'_out.log', 'w'),
open(
os.path.join(voxel_globe.tools.log_dir(), self.request.id) +
'_err.log', 'w')):
openclDevice = os.environ['VIP_OPENCL_DEVICE']
opencl_memory = os.environ.get('VIP_OPENCL_MEMORY', None)
if opencl_memory:
opencl_memory = int(opencl_memory)
scene = models.Scene.objects.get(id=scene_id)
imageCollection = models.ImageCollection.objects.get(\
id=image_collection_id).history(history)
imageList = imageCollection.images.all()
with voxel_globe.tools.task_dir('voxel_world') as processing_dir:
logger.warning(bbox)
if bbox['geolocated']:
create_scene_xml(
openclDevice,
3,
float(bbox['voxel_size']),
lla1=(float(bbox['x_min']), float(bbox['y_min']),
float(bbox['z_min'])),
lla2=(float(bbox['x_max']), float(bbox['y_max']),
float(bbox['z_max'])),
origin=scene.origin,
model_dir='.',
number_bins=1,
output_file=open(os.path.join(processing_dir, 'scene.xml'),
'w'),
n_bytes_gpu=opencl_memory)
else:
create_scene_xml(
openclDevice,
3,
float(bbox['voxel_size']),
lvcs1=(float(bbox['x_min']), float(bbox['y_min']),
float(bbox['z_min'])),
lvcs2=(float(bbox['x_max']), float(bbox['y_max']),
float(bbox['z_max'])),
origin=scene.origin,
model_dir='.',
number_bins=1,
output_file=open(os.path.join(processing_dir, 'scene.xml'),
'w'),
n_bytes_gpu=opencl_memory)
counter = 1
imageNames = []
cameraNames = []
os.mkdir(os.path.join(processing_dir, 'local'))
#Prepping
for image in imageList:
self.update_state(state='INITIALIZE',
meta={
'stage': 'image fetch',
'i': counter,
'total': len(imageList)
})
image = image.history(history)
(K, R, T, o) = get_krt(image.history(history), history=history)
krtName = os.path.join(processing_dir, 'local',
'frame_%05d.krt' % counter)
with open(krtName, 'w') as fid:
print >> fid, (("%0.18f " * 3 + "\n") * 3) % (
K[0, 0], K[0, 1], K[0, 2], K[1, 0], K[1, 1], K[1, 2],
K[2, 0], K[2, 1], K[2, 2])
print >> fid, (("%0.18f " * 3 + "\n") * 3) % (
R[0, 0], R[0, 1], R[0, 2], R[1, 0], R[1, 1], R[1, 2],
R[2, 0], R[2, 1], R[2, 2])
print >> fid, ("%0.18f " * 3 + "\n") % (T[0, 0], T[1, 0],
T[2, 0])
imageName = image.originalImageUrl
extension = os.path.splitext(imageName)[1]
localName = os.path.join(processing_dir, 'local',
'frame_%05d%s' % (counter, extension))
wget(imageName, localName, secret=True)
counter += 1
imageNames.append(localName)
cameraNames.append(krtName)
variance = 0.06
vxl_scene = boxm2_scene_adaptor(
os.path.join(processing_dir, "scene.xml"), openclDevice)
current_level = 0
loaded_imgs = []
loaded_cams = []
for i in range(0, len(imageNames), skip_frames):
logger.debug("i: %d img name: %s cam name: %s", i,
imageNames[i], cameraNames[i])
self.update_state(state='PRELOADING',
meta={
'stage': 'image load',
'i': i,
'total': len(imageNames)
})
img, ni, nj = load_image(imageNames[i])
loaded_imgs.append(img)
pcam = load_perspective_camera(cameraNames[i])
loaded_cams.append(pcam)
refine_cnt = 5
refine_device = openclDevice[0:3]
if refine_device == 'cpu':
refine_device = 'cpp'
for rfk in range(0, refine_cnt, 1):
pair = zip(loaded_imgs, loaded_cams)
random.shuffle(pair)
for idx, (img, cam) in enumerate(pair):
self.update_state(state='PROCESSING',
meta={
'stage': 'update',
'i': rfk + 1,
'total': refine_cnt,
'image': idx + 1,
'images': len(loaded_imgs)
})
logger.debug("refine_cnt: %d, idx: %d", rfk, idx)
vxl_scene.update(cam,
img,
True,
True,
None,
openclDevice[0:3],
variance,
tnear=1000.0,
tfar=100000.0)
for bp_index in range(2):
pass #height map update?
logger.debug("writing cache: %d", rfk)
vxl_scene.write_cache()
logger.debug("wrote cache: %d", rfk)
if rfk < refine_cnt - 1:
self.update_state(state='PROCESSING',
meta={
'stage': 'refine',
'i': rfk,
'total': refine_cnt
})
logger.debug("refining %d...", rfk)
vxl_scene.refine(0.3, refine_device)
vxl_scene.write_cache()
with open(os.path.join(processing_dir, "scene_color.xml"),
'w') as fid:
lines = open(os.path.join(processing_dir, "scene.xml"),
'r').readlines()
lines = [
line.replace('boxm2_mog3_grey', 'boxm2_gauss_rgb').replace(
'boxm2_num_obs', 'boxm2_num_obs_single')
for line in lines
]
fid.writelines(lines)
vxl_scene = boxm2_scene_adaptor(
os.path.join(processing_dir, "scene_color.xml"), openclDevice)
for idx, (img, cam) in enumerate(pair):
self.update_state(state='PROCESSING',
meta={
'stage': 'color_update',
'i': rfk + 1,
'total': refine_cnt,
'image': idx + 1,
'images': len(loaded_imgs)
})
logger.debug("color_paint idx: %d", idx)
vxl_scene.update(cam,
img,
False,
False,
None,
openclDevice[0:3],
tnear=1000.0,
tfar=100000.0)
vxl_scene.write_cache()
with voxel_globe.tools.storage_dir(
'voxel_world') as voxel_world_dir:
copytree(processing_dir,
voxel_world_dir,
ignore=lambda x, y: ['images'])
models.VoxelWorld.create(
name='%s world (%s)' %
(imageCollection.name, self.request.id),
origin=scene.origin,
directory=voxel_world_dir,
service_id=self.request.id).save()
def generate_error_point_cloud(
self, voxel_world_id, prob=0.5, position_error_override=None, orientation_error_override=None, history=None
):
from glob import glob
import json
import numpy as np
from boxm2_adaptor import (
load_cpp,
render_depth,
cast_3d_point,
cast_3d_point_pass2,
write_cache,
accumulate_3d_point_and_cov,
normalize_3d_point_and_cov,
)
from boxm2_scene_adaptor import boxm2_scene_adaptor
from vpgl_adaptor import create_perspective_camera_krt, persp2gen, compute_direction_covariance
from boxm2_mesh_adaptor import gen_error_point_cloud
from vsi.tools.redirect import StdRedirect, Logger as LoggerWrapper
import voxel_globe.tools
import voxel_globe.meta.models as models
from voxel_globe.tools.camera import get_krt
with StdRedirect(
open(os.path.join(voxel_globe.tools.log_dir(), self.request.id) + "_out.log", "w"),
open(os.path.join(voxel_globe.tools.log_dir(), self.request.id) + "_err.log", "w"),
):
self.update_state(state="SETUP", meta={"pid": os.getpid()})
voxel_world = models.VoxelWorld.objects.get(id=voxel_world_id)
service_inputs = json.loads(voxel_world.service.inputs)
image_collection = models.ImageCollection.objects.get(id=service_inputs[0][0])
images = image_collection.images.all()
scene = models.Scene.objects.get(id=service_inputs[0][1])
voxel_world_dir = voxel_world.directory
scene_filename = os.path.join(voxel_world_dir, "scene_color.xml")
opencl_device = os.environ["VIP_OPENCL_DEVICE"]
scene_gpu = boxm2_scene_adaptor(scene_filename, opencl_device)
scene_cpp = boxm2_scene_adaptor(scene_filename, "cpp")
type_id_fname = "type_names_list.txt"
image_id_fname = "image_list.txt"
std_dev_angle_default = 0.1
cov_c_path = "cov_c.txt"
cov_c_default = 0.8
with voxel_globe.tools.task_dir("generate_error_point_cloud", cd=True) as processing_dir:
for index, image in enumerate(images):
self.update_state(
state="PROCESSING", meta={"stage": "casting", "image": index + 1, "total": len(images)}
)
k, r, t, o = get_krt(image.history(history))
attributes = image.history(history).camera.history(history).attributes
cov_c = attributes.get("position_error", std_dev_angle_default)
if position_error_override is not None:
cov_c = position_error_override
std_dev_angle = attributes.get("orientation_error", cov_c_default)
if orientation_error_override is not None:
std_dev_angle = orientation_error_override
cov_c = np.eye(3) * cov_c ** 2
np.savetxt(cov_c_path, cov_c)
perspective_camera = create_perspective_camera_krt(k, r, t)
self.update_state(
state="PROCESSING", meta={"stage": "pre render", "image": index + 1, "total": len(images)}
)
(depth_image, variance_image, _) = render_depth(
scene_gpu.scene,
scene_gpu.opencl_cache,
perspective_camera,
image.imageWidth,
image.imageHeight,
scene_gpu.device,
)
self.update_state(
state="PROCESSING", meta={"stage": "post_render", "image": index + 1, "total": len(images)}
)
cov_v_path = "cov_%06d.txt" % index
appearance_model = "image"
self.update_state(
state="PROCESSING", meta={"stage": "pre_persp2gen", "image": index + 1, "total": len(images)}
)
generic_camera = persp2gen(perspective_camera, image.imageWidth, image.imageHeight)
self.update_state(
state="PROCESSING", meta={"stage": "pre_covar", "image": index + 1, "total": len(images)}
)
compute_direction_covariance(perspective_camera, std_dev_angle, cov_v_path)
self.update_state(
state="PROCESSING", meta={"stage": "pre_cast1", "image": index + 1, "total": len(images)}
)
cast_3d_point(
scene_cpp.scene,
scene_cpp.cpu_cache,
perspective_camera,
generic_camera,
depth_image,
variance_image,
appearance_model,
)
self.update_state(
state="PROCESSING", meta={"stage": "pre_cast2", "image": index + 1, "total": len(images)}
)
cast_3d_point_pass2(
scene_cpp.scene, scene_cpp.cpu_cache, generic_camera, appearance_model, cov_c_path, cov_v_path
)
self.update_state(
state="PROCESSING", meta={"stage": "pre_accumulate", "image": index + 1, "total": len(images)}
)
accumulate_3d_point_and_cov(scene_cpp.scene, scene_cpp.cpu_cache, appearance_model)
# self.update_state(state='PROCESSING',
# meta={'stage':'pre_write', 'image':index+1,
# 'total':len(images)})
# write_cache(scene_cpp.cpu_cache, 1)
self.update_state(
state="PROCESSING", meta={"stage": "post_write", "image": index + 1, "total": len(images)}
)
self.update_state(state="PROCESSING", meta={"stage": "compute error"})
normalize_3d_point_and_cov(scene_cpp.scene, scene_cpp.cpu_cache)
self.update_state(state="PROCESSING", meta={"stage": "pre_write", "image": index + 1, "total": len(images)})
write_cache(scene_cpp.cpu_cache, 1)
self.update_state(state="EXPORTING", meta={"stage": "ply"})
with voxel_globe.tools.storage_dir("generate_error_point_cloud") as storage_dir:
ply_filename = os.path.join(storage_dir, "model.ply")
gen_error_point_cloud(scene_cpp.scene, scene_cpp.cpu_cache, ply_filename, prob)
potree_filename = os.path.join(storage_dir, "potree.ply")
with voxel_globe.tools.image_dir("point_cloud") as potree_dir:
convert_ply_to_potree(ply_filename, potree_dir)
models.PointCloud.create(
name="%s point cloud" % image_collection.name,
service_id=self.request.id,
origin=voxel_world.origin,
potree_url="%s://%s:%s/%s/point_cloud/%s/cloud.js"
% (
env["VIP_IMAGE_SERVER_PROTOCOL"],
env["VIP_IMAGE_SERVER_HOST"],
env["VIP_IMAGE_SERVER_PORT"],
env["VIP_IMAGE_SERVER_URL_PATH"],
os.path.basename(potree_dir),
),
filename=ply_filename,
).save()
voxel_files = lambda x: glob(os.path.join(voxel_world_dir, x))
cleanup_files = []
cleanup_files += voxel_files("boxm2_covariance_*.bin")
cleanup_files += voxel_files("boxm2_point_*.bin")
cleanup_files += voxel_files("float16_image_*.bin")
for cleanup_file in cleanup_files:
os.remove(cleanup_file)
def run_build_voxel_model_bp(self,
image_set_id,
camera_set_id,
scene_id,
bbox,
skip_frames,
cleanup=True):
import random
import glob
import math
import numpy as np
from vsi.tools.redirect import StdRedirect
from voxel_globe.meta import models
from voxel_globe.tools.camera import get_krt
import voxel_globe.tools
from boxm2_scene_adaptor import boxm2_scene_adaptor
import brl_init
from vil_adaptor_boxm2_batch import load_image
import vpgl_adaptor_boxm2_batch as vpgl
from vsi.vxl.create_scene_xml import create_scene_xml
from vsi.tools.dir_util import copytree
from vsi.tools.file_util import lncp
def rectint(recta, rectb):
lx = max(recta[0], rectb[0])
rx = min(recta[2], rectb[2])
by = max(recta[1], rectb[1])
ty = min(recta[3], rectb[3])
if lx > rx or by > ty:
return [0, 0, 0, 0], 0
else:
return [lx, by, rx, ty], (rx - lx) * (ty - by)
def generate_subsetim(scene, camfiles, ni, nj):
subsetIdx = []
refIndices = []
minDepOverlap = 0.25
minRefOverlap = 0.5
minIndepAngle = 5.0
minRefAngle = 5.0
maxRefAngle = 15.0
minRefIndepAngle = 5.0
cosMinIndepAngle = math.cos(minIndepAngle * math.pi / 180.0)
cosMinRefAngle = math.cos(minRefAngle * math.pi / 180.0)
cosMaxRefAngle = math.cos(maxRefAngle * math.pi / 180.0)
cosMinRefIndepAngle = math.cos(minRefIndepAngle * math.pi / 180.0)
bbox = scene.bbox
grect = [
scene.bbox[0][0], scene.bbox[0][1], scene.bbox[1][0],
scene.bbox[1][1]
]
worldoverlaps = []
camrects = []
cams = []
princAxis = []
for camfile in camfiles:
pcam = vpgl.load_perspective_camera(camfile)
prx, pry, prz = vpgl.get_backprojected_ray(pcam, ni / 2, nj / 2)
princAxis.append([prx, pry, prz])
Hmat = vpgl.compute_camera_to_world_homography(
pcam, [0, 0, 1, -bbox[0][2]])
H = np.array(Hmat).reshape([3, 3])
ps = np.dot(
H,
np.transpose([[0, 0, 1], [ni, 0, 1], [ni, nj, 1], [0, nj, 1]]))
xs = ps[0, :] / ps[2, :]
ys = ps[1, :] / ps[2, :]
rect = [min(xs), min(ys), max(xs), max(ys)]
area = (rect[2] - rect[0]) * (rect[3] - rect[1])
crect, carea = rectint(rect, grect)
#print crect,carea
if (carea > 0):
cams.append(pcam)
camrects.append(crect)
worldoverlaps.append(carea / area)
usedcams = [False] * len(cams)
for i in range(0, len(cams)):
randidx = random.randint(0, len(cams) - 1)
while usedcams[randidx]:
randidx = (randidx + 1) % len(cams)
usedcams[randidx] = True
dep = False
for c2 in range(0, len(subsetIdx)):
cosAngle = np.dot(princAxis[randidx], princAxis[subsetIdx[c2]])
if cosAngle > cosMinIndepAngle:
rectc2 = camrects[subsetIdx[c2]]
overlap, oarea = rectint(camrects[randidx], rectc2)
tarea = (rectc2[2] - rectc2[0]) * (rectc2[3] - rectc2[1])
if (oarea / tarea > minDepOverlap):
dep = True
break
if dep:
continue
theseRefIndices = []
for c3 in range(0, len(cams)):
#Check angle disparity
cosAngle2 = np.dot(princAxis[randidx], princAxis[c3])
if (cosAngle2 > cosMinRefAngle or cosAngle2 < cosMaxRefAngle):
continue
# Check that a similar viewpoint isn't already used for reference
refDep = False
for c4 in range(0, len(theseRefIndices)):
#Check angle disparity
cosAngle3 = np.dot(princAxis[theseRefIndices[c4]],
princAxis[c3])
if (cosAngle3 > cosMinRefIndepAngle):
refDep = True
break
#If similar viewpoint don't add
if (refDep):
continue
theseRefIndices.append(c3)
#If at least one reference image save this viewpoint
if len(theseRefIndices) > 0:
subsetIdx.append(randidx)
refIndices.append(theseRefIndices)
return subsetIdx, refIndices
def update_bp(scene,
images,
cameras,
do_update_image=True,
do_update_hmap=False):
_, ni, nj = load_image(images[0])
frames, refimages = generate_subsetim(scene, cameras, ni, nj)
for file_name in glob.glob(os.path.join(scene.model_dir,
'boxm2_*.bin')):
os.remove(file_name)
scene.init_uniform_prob()
sradius = 16
idents = []
weights = []
if do_update_image:
idents.append("if")
weights.append(1.0)
if do_update_hmap:
idents.append("hf")
weights.append(2.0)
for idx, i in enumerate(frames):
if do_update_image:
print "Iteration ", idx, "Image ", images[i]
####load image and camera
viewid = os.path.splitext(os.path.basename(images[i]))[0]
#### forming an app model using the neighbor images
for lindex in refimages[idx]:
lcam = vpgl.load_perspective_camera(cameras[lindex])
limg, ni, nj = load_image(images[lindex])
scene.update(lcam, limg, False, True, None, "gpu0", 0.05,
viewid)
scene.update_if(False, viewid) # subtracting the image factor
scene.fuse_factors(idents, weights)
pcam = vpgl.load_perspective_camera(cameras[i])
img, ni, nj = load_image(images[i])
scene.compute_pre_post(pcam, img, viewid, 100000, 100000)
# computing the new image factor
scene.update_if(True, viewid) # adding the image factor
scene.fuse_factors(idents, weights)
if do_update_hmap and idx % 2 == 0:
scene.update_hf(False) # subtracting the height-map factor
scene.fuse_factors(idents, weights)
zimg, zvar, ximg, yimg, probimg = scene.render_height_map()
#save_image(zimg, "./zimg.tif")
scene.compute_hmapf(zimg, zvar, ximg, yimg,
sradius) # computing the height-map factor
scene.update_hf(True) # adding the height-map factor
scene.fuse_factors(idents, weights)
scene.write_cache()
def refine(scene):
scene.refine(0.3)
for filename in glob.glob(os.path.join(scene.model_dir, '[a-b]*.bin')):
os.remove(filename)
scene.write_cache()
with StdRedirect(
open(
os.path.join(voxel_globe.tools.log_dir(), self.request.id) +
'_out.log', 'w'),
open(
os.path.join(voxel_globe.tools.log_dir(), self.request.id) +
'_err.log', 'w')):
openclDevice = os.environ['VIP_OPENCL_DEVICE']
opencl_memory = os.environ.get('VIP_OPENCL_MEMORY', None)
if opencl_memory:
opencl_memory = int(opencl_memory)
scene = models.Scene.objects.get(id=scene_id)
imageSet = models.ImageSet.objects.get(id=image_set_id)
imageList = imageSet.images.all()
with voxel_globe.tools.task_dir('voxel_world') as processing_dir:
logger.warning(bbox)
create_scene_xml(openclDevice,
0,
float(bbox['voxel_size']),
lvcs1=(float(bbox['x_min']), float(bbox['y_min']),
float(bbox['z_min'])),
lvcs2=(float(bbox['x_max']), float(bbox['y_max']),
float(bbox['z_max'])),
origin=scene.origin,
model_dir='.',
number_bins=1,
output_file=open(
os.path.join(processing_dir, 'scene.xml'),
'w'),
n_bytes_gpu=opencl_memory)
counter = 1
imageNames = []
cameraNames = []
os.mkdir(os.path.join(processing_dir, 'local'))
#Prepping
self.update_state(state='INITIALIZE',
meta={
'image_set_name': imageSet.name,
'stage': 'camera fetch'
})
for image in imageList:
(K, R, T, o) = get_krt(image, camera_set_id)
krtName = os.path.join(processing_dir, 'local',
'frame_%05d.krt' % counter)
with open(krtName, 'w') as fid:
print >> fid, (("%0.18f " * 3 + "\n") * 3) % (
K[0, 0], K[0, 1], K[0, 2], K[1, 0], K[1, 1], K[1, 2],
K[2, 0], K[2, 1], K[2, 2])
print >> fid, (("%0.18f " * 3 + "\n") * 3) % (
R[0, 0], R[0, 1], R[0, 2], R[1, 0], R[1, 1], R[1, 2],
R[2, 0], R[2, 1], R[2, 2])
print >> fid, ("%0.18f " * 3 + "\n") % (T[0, 0], T[1, 0],
T[2, 0])
imageName = image.filename_path
extension = os.path.splitext(imageName)[1]
localName = os.path.join(processing_dir, 'local',
'frame_%05d%s' % (counter, extension))
lncp(imageName, localName)
counter += 1
imageNames.append(localName)
cameraNames.append(krtName)
variance = 0.06
vxl_scene = boxm2_scene_adaptor(
os.path.join(processing_dir, "scene.xml"), openclDevice)
# loaded_imgs = []
# loaded_cams = []
# for i in range(0, len(imageNames), skip_frames):
# logger.debug("i: %d img name: %s cam name: %s", i, imageNames[i],
# cameraNames[i])
# self.update_state(state='PRELOADING', meta={'image_set_name': imageSet.name,
# 'stage':'image load',
# 'i':i,
# 'total':len(imageNames)})
# img, ni, nj = load_image(imageNames[i])
# loaded_imgs.append(img)
# pcam = load_perspective_camera(cameraNames[i])
# loaded_cams.append(pcam)
refine_cnt = 2
for rfk in range(0, refine_cnt, 1):
self.update_state(state='PROCESSING',
meta={
'image_set_name': imageSet.name,
'stage': 'update 1'
})
update_bp(vxl_scene, imageNames, cameraNames)
# self.update_state(state='PROCESSING', meta={'image_set_name': imageSet.name,
# 'stage':'update 2'})
# update_bp(vxl_scene, imageNames, cameraNames, True, True)
# self.update_state(state='PROCESSING', meta={'image_set_name': imageSet.name,
# 'stage':'update 3'})
# update_bp(vxl_scene, imageNames, cameraNames, True, True)
if rfk < refine_cnt - 1:
self.update_state(state='PROCESSING',
meta={
'image_set_name': imageSet.name,
'stage': 'refine',
'i': rfk + 1,
'total': refine_cnt
})
refine(vxl_scene)
#Update color appearance
with open(os.path.join(processing_dir, "scene_color.xml"),
'w') as fid:
lines = open(os.path.join(processing_dir, "scene.xml"),
'r').readlines()
lines = [
line.replace('boxm2_mog3_grey', 'boxm2_gauss_rgb').replace(
'boxm2_num_obs', 'boxm2_num_obs_single')
for line in lines
]
fid.writelines(lines)
vxl_scene = boxm2_scene_adaptor(
os.path.join(processing_dir, "scene_color.xml"), openclDevice)
for idx, (image_name,
camera_name) in enumerate(zip(imageNames, cameraNames)):
self.update_state(state='PROCESSING',
meta={
'image_set_name': imageSet.name,
'stage': 'color_update',
'i': idx + 1,
'total': len(imageNames),
'images': len(imageNames)
})
img, _, _ = load_image(image_name)
pcam = vpgl.load_perspective_camera(camera_name)
logger.debug("color_paint idx: %d", idx)
vxl_scene.update(pcam,
img,
False,
False,
None,
openclDevice,
tnear=1000.0,
tfar=100000.0)
vxl_scene.write_cache()
with voxel_globe.tools.storage_dir(
'voxel_world') as voxel_world_dir:
copytree(processing_dir,
voxel_world_dir,
ignore=lambda x, y: ['local'])
models.VoxelWorld(name='%s world (%s)' %
(imageSet.name, self.request.id),
origin=scene.origin,
directory=voxel_world_dir,
service_id=self.request.id).save()
return {"image_set_name": imageSet.name}
def runVisualSfm(self, imageSetId, sceneId, cleanup=True):
from voxel_globe.meta import models
from os import environ as env
from os.path import join as path_join
import os
import shutil
import time
from django.contrib.gis.geos import Point
from .tools import writeNvm, writeGcpFile, generateMatchPoints, runSparse,\
readNvm
import voxel_globe.tools
from voxel_globe.tools.camera import get_kto, save_krt
import voxel_globe.tools.enu as enu
import numpy
import boxm2_adaptor
import boxm2_scene_adaptor
from voxel_globe.tools.xml_dict import load_xml
from django.contrib.gis.geos import Point
from voxel_globe.tools.image import convert_image
from distutils.spawn import find_executable
from vsi.iglob import glob as glob
self.update_state(state='INITIALIZE', meta={'stage':0})
#Make main temp dir and cd into it
with voxel_globe.tools.task_dir('visualsfm', cd=True) as processing_dir:
#Because visualsfm is so... bad, I need to copy it locally so I can
#configure it
visualsfm_exe = os.path.join(processing_dir, 'visualsfm')
shutil.copy(find_executable('VisualSFM'), visualsfm_exe)
with open(os.path.join(processing_dir, 'nv.ini'), 'w') as fid:
fid.write('param_search_multiple_models 0\n')
fid.write('param_use_siftgpu 2\n')
matchFilename = path_join(processing_dir, 'match.nvm')
sparce_filename = path_join(processing_dir, 'sparse.nvm')
#This can NOT be changed in version 0.5.25
gcpFilename = matchFilename + '.gcp'
logger.debug('Task %s is processing in %s' % (self.request.id,
processing_dir))
image_set = models.ImageSet.objects.get(
id=imageSetId)
imageList = image_set.images.all()
### if 1:
### try: #Not fully integrated yet
### sift_gpu = siftgpu.SiftGPU()
### except:
### pass
localImageList = []
for x in range(len(imageList)):
#Download the image locally
image = imageList[x]
self.update_state(state='INITIALIZE', meta={'stage':'image fetch', 'i':x,
'total':len(imageList)})
imageName = image.filename_path
extension = os.path.splitext(imageName)[1].lower()
localName = path_join(processing_dir, 'frame_%05d%s' % (x+1, extension))
#lncp(imageName, localName)
#Stupid VisualSFM dereferences symlinks, breaking this
shutil.copyfile(imageName, localName)
#Convert the image if necessary
if extension not in ['.jpg', '.jpeg', '.pgm', '.ppm']:
self.update_state(state='INITIALIZE',
meta={'stage':'image convert', 'i':x, 'total':len(imageList)})
#Add code here to converty to jpg for visual sfm
if extension in ['.png']:#'not implemented':
from PIL import Image
image_temp = Image.open(localName)
# if len(image_temp.mode) > 1: #Stupid visual sfm is picky :(
# new_local_name = os.path.splitext(localName)[0] + '.ppm'
# else:
# new_local_name = os.path.splitext(localName)[0] + '.pgm'
new_local_name = os.path.splitext(localName)[0] + '.jpg'
###ingest.convert_image(localName, new_local_name, 'PNM')
convert_image(localName, new_local_name, 'JPEG',
options=('QUALITY=100',))
os.remove(localName)
localName = new_local_name
else:
raise Exception('Unsupported file type')
imageInfo = {'localName':localName, 'index':x}
try:
[K, T, llh] = get_kto(image)
imageInfo['K_intrinsics'] = K
imageInfo['transformation'] = T
imageInfo['enu_origin'] = llh
except:
pass
localImageList.append(imageInfo)
### if 1:
### try: #not fully integrated yet
### sift_gpu.create_sift(localName, os.path.splitext(localName)[0]+'.sift')
### except:
### pass
# filenames = list(imageList.values_list('image_url'))
# logger.info('The image list 0is %s' % filenames)
self.update_state(state='PROCESSING',
meta={'stage':'generate match points',
'processing_dir':processing_dir,
'total':len(imageList)})
pid = generateMatchPoints(map(lambda x:x['localName'], localImageList),
matchFilename, logger=logger,
executable=visualsfm_exe)
old_mat=None
old_sift=None
#TODO: Replace with inotify to monitor directory
while pid.poll() is None:
mat = len(glob(os.path.join(processing_dir, '*.mat'), False))
sift = len(glob(os.path.join(processing_dir, '*.sift'), False))
if mat != old_mat or \
sift != old_sift:
old_mat=mat
old_sift=sift
self.update_state(state='PROCESSING',
meta={'stage':'generate match points',
'processing_dir':processing_dir,
'sift':sift,
'mat':mat,
'total':len(imageList)})
time.sleep(5)
# cameras = []
# for image in imageList:
# if 1:
# #try:
# [K, T, llh] = get_kto(image)
# cameras.append({'image':image.id, 'K':K, 'tranformation':
# T, 'origin':llh})
# #except:
# pass
# origin = numpy.median(origin, axis=0)
# origin = [-92.215197, 37.648858, 268.599]
scene = models.Scene.objects.get(id=sceneId)
origin = list(scene.origin)
if scene.geolocated:
self.update_state(state='PROCESSING',
meta={'stage':'writing gcp points'})
#find the middle origin, and make it THE origin
data = []#.name .llh_xyz
for imageInfo in localImageList:
try:
r = imageInfo['transformation'][0:3, 0:3]
t = imageInfo['transformation'][0:3, 3:]
enu_point = -r.transpose().dot(t)
if not numpy.array_equal(imageInfo['enu_origin'], origin):
ecef = enu.enu2xyz(refLong=imageInfo['enu_origin'][0],
refLat=imageInfo['enu_origin'][1],
refH=imageInfo['enu_origin'][2],
#e=imageInfo['transformation'][0, 3],
#n=imageInfo['transformation'][1, 3],
#u=imageInfo['transformation'][2, 3])
e=enu_point[0],
n=enu_point[1],
u=enu_point[2])
enu_point = enu.xyz2enu(refLong=origin[0],
refLat=origin[1],
refH=origin[2],
X=ecef[0],
Y=ecef[1],
Z=ecef[2])
# else:
# enu_point = imageInfo['transformation'][0:3, 3]
dataBit = {'filename':imageInfo['localName'], 'xyz':enu_point}
data.append(dataBit)
#Make this a separate ingest process, making CAMERAS linked to the
#images
#data = arducopter.loadAdjTaggedMetadata(
# r'd:\visualsfm\2014-03-20 13-22-44_adj_tagged_images.txt')
#Make this read the cameras from the DB instead
writeGcpFile(data, gcpFilename)
except: #some images may have no camera
pass
self.update_state(state='PROCESSING', meta={'stage':'sparse SFM'})
pid = runSparse(matchFilename, sparce_filename, gcp=scene.geolocated,
shared=True, logger=logger, executable=visualsfm_exe)
pid.wait()
self.update_state(state='FINALIZE',
meta={'stage':'loading resulting cameras'})
#prevent bundle2scene from getting confused and crashing
sift_data = os.path.join(processing_dir, 'sift_data')
os.mkdir(sift_data)
for filename in glob(os.path.join(processing_dir, '*.mat'), False) +\
glob(os.path.join(processing_dir, '*.sift'), False):
shutil.move(filename, sift_data)
if scene.geolocated:
#Create a uscene.xml for the geolocated case. All I want out of this is
#the bounding box and gsd calculation.
boxm2_adaptor.bundle2scene(sparce_filename, processing_dir, isalign=False,
out_dir="")
cams = readNvm(path_join(processing_dir, 'sparse.nvm'))
#cams.sort(key=lambda x:x.name)
#Since the file names are frame_00001, etc... and you KNOW this order is
#identical to localImageList, with some missing
camera_set = models.CameraSet(name="Visual SFM Geo %s" % image_set.name,
service_id = self.request.id,
images_id = imageSetId)
camera_set.save()
for cam in cams:
frameName = cam.name #frame_00001, etc....
imageInfo = filter(lambda x: x['localName'].endswith(frameName),
localImageList)[0]
#I have to use endswith instead of == because visual sfm APPARENTLY
#decides to take some liberty and make absolute paths relative
image = imageList[imageInfo['index']]
(k,r,t) = cam.krt(width=image.image_width, height=image.image_height)
t = t.flatten()
camera = save_krt(self.request.id, image, k, r, t, origin, srid=4326)
camera_set.cameras.add(camera)
else:
from vsi.tools.natural_sort import natural_sorted
from vsi.io.krt import Krt
boxm2_adaptor.bundle2scene(sparce_filename, processing_dir, isalign=True,
out_dir=processing_dir)
#While the output dir is used for the b2s folders, uscene.xml is cwd
#They are both set to processing_dir, so everything works out well
aligned_cams = glob(os.path.join(processing_dir, 'cams_krt', '*'))
#sort them naturally in case there are more then 99,999 files
aligned_cams = natural_sorted(aligned_cams)
if len(aligned_cams) != len(imageList):
#Create a new image set
new_image_set = models.ImageSet(
name="SFM Result Subset (%s)" % image_set.name,
service_id = self.request.id)
# for image in image_set.images.all():
# new_image_set.images.add(image)
new_image_set.save()
frames_keep = set(map(lambda x:
int(os.path.splitext(x.split('_')[-2])[0])-1, aligned_cams))
for frame_index in frames_keep:
new_image_set.images.add(imageList[frame_index])
# frames_remove = set(xrange(len(imageList))) - frames_keep
#
# for remove_index in list(frames_remove):
# #The frame number refers to the nth image in the image set,
# #so frame_00100.tif is the 100th image, starting the index at one
# #See local_name above
#
# #remove the images sfm threw away
# new_image_set.remove(imageList[remove_index])
image_set = new_image_set
frames_keep = list(frames_keep)
else:
frames_keep = xrange(len(aligned_cams))
camera_set = models.CameraSet(name="Visual SFM %s" % image_set.name,
service_id = self.request.id,
images_id = imageSetId)
camera_set.save()
#---Update the camera models in the database.---
for camera_index, frame_index in enumerate(frames_keep):
krt = Krt.load(aligned_cams[camera_index])
image = imageList[frame_index]
camera = save_krt(self.request.id, image, krt.k, krt.r, krt.t, [0,0,0],
srid=4326)
camera_set.cameras.add(camera)
#---Update scene information important for the no-metadata case ---
scene_filename = os.path.join(processing_dir, 'model', 'uscene.xml')
boxm_scene = boxm2_scene_adaptor.boxm2_scene_adaptor(scene_filename)
scene.bbox_min = Point(*boxm_scene.bbox[0])
scene.bbox_max = Point(*boxm_scene.bbox[1])
#This is not a complete or good function really... but it will get me the
#information I need.
scene_dict = load_xml(scene_filename)
block = scene_dict['block']
scene.default_voxel_size=Point(float(block.at['dim_x']),
float(block.at['dim_y']),
float(block.at['dim_z']))
scene.save()
def generate_error_point_cloud(self,
voxel_world_id,
camera_set_id,
prob=0.5,
position_error_override=None,
orientation_error_override=None,
number_images=None):
from glob import glob
import json
import random
import numpy as np
from boxm2_adaptor import load_cpp, render_depth, cast_3d_point, \
cast_3d_point_pass2, write_cache, \
accumulate_3d_point_and_cov, \
normalize_3d_point_and_cov
from boxm2_scene_adaptor import boxm2_scene_adaptor
from vpgl_adaptor_boxm2_batch import create_perspective_camera_krt, \
persp2gen, compute_direction_covariance
from boxm2_mesh_adaptor import gen_error_point_cloud
from vsi.tools.redirect import StdRedirect, Logger as LoggerWrapper
import voxel_globe.tools
import voxel_globe.meta.models as models
from voxel_globe.tools.camera import get_krt
with StdRedirect(
open(
os.path.join(voxel_globe.tools.log_dir(), self.request.id) +
'_out.log', 'w'),
open(
os.path.join(voxel_globe.tools.log_dir(), self.request.id) +
'_err.log', 'w')):
self.update_state(state='SETUP', meta={'pid': os.getpid()})
voxel_world = models.VoxelWorld.objects.get(id=voxel_world_id)
service_inputs = json.loads(voxel_world.service.inputs)
image_set = models.ImageSet.objects.get(id=service_inputs[0][0])
images = image_set.images.all()
voxel_world_dir = voxel_world.directory
scene_filename = os.path.join(voxel_world_dir, 'scene_color.xml')
opencl_device = os.environ['VIP_OPENCL_DEVICE']
scene_gpu = boxm2_scene_adaptor(scene_filename, opencl_device)
scene_cpp = boxm2_scene_adaptor(scene_filename, 'cpp')
type_id_fname = "type_names_list.txt"
image_id_fname = "image_list.txt"
std_dev_angle_default = 0
cov_c_path = 'cov_c.txt'
cov_c_default = 0
if not number_images:
number_images = len(images)
number_images = min(len(images), number_images)
with voxel_globe.tools.task_dir('generate_error_point_cloud', cd=True) \
as processing_dir:
for index, image in enumerate(random.sample(images,
number_images)):
self.update_state(state='PROCESSING',
meta={
'stage': 'casting',
'image': index + 1,
'total': len(images)
})
k, r, t, o = get_krt(image, camera_set_id)
attributes = image.camera_set.get(
cameraset=camera_set_id).attributes
cov_c = attributes.get('position_error', cov_c_default)
if position_error_override is not None:
cov_c = position_error_override
std_dev_angle = attributes.get('orientation_error',
std_dev_angle_default)
if orientation_error_override is not None:
std_dev_angle = orientation_error_override
cov_c = np.eye(3) * cov_c**2
np.savetxt(cov_c_path, cov_c)
perspective_camera = create_perspective_camera_krt(k, r, t)
self.update_state(state='PROCESSING',
meta={
'stage': 'pre render',
'image': index + 1,
'total': len(images)
})
(depth_image, variance_image,
_) = render_depth(scene_gpu.scene, scene_gpu.opencl_cache,
perspective_camera, image.image_width,
image.image_height, scene_gpu.device)
self.update_state(state='PROCESSING',
meta={
'stage': 'post_render',
'image': index + 1,
'total': len(images)
})
cov_v_path = 'cov_%06d.txt' % index
appearance_model = 'image'
self.update_state(state='PROCESSING',
meta={
'stage': 'pre_persp2gen',
'image': index + 1,
'total': len(images)
})
generic_camera = persp2gen(perspective_camera,
image.image_width,
image.image_height)
self.update_state(state='PROCESSING',
meta={
'stage': 'pre_covar',
'image': index + 1,
'total': len(images)
})
compute_direction_covariance(perspective_camera, std_dev_angle,
cov_v_path)
self.update_state(state='PROCESSING',
meta={
'stage': 'pre_cast1',
'image': index + 1,
'total': len(images)
})
cast_3d_point(scene_cpp.scene, scene_cpp.cpu_cache,
perspective_camera, generic_camera, depth_image,
variance_image, appearance_model)
self.update_state(state='PROCESSING',
meta={
'stage': 'pre_cast2',
'image': index + 1,
'total': len(images)
})
cast_3d_point_pass2(scene_cpp.scene, scene_cpp.cpu_cache,
generic_camera, appearance_model,
cov_c_path, cov_v_path)
self.update_state(state='PROCESSING',
meta={
'stage': 'pre_accumulate',
'image': index + 1,
'total': len(images)
})
accumulate_3d_point_and_cov(scene_cpp.scene,
scene_cpp.cpu_cache,
appearance_model)
#self.update_state(state='PROCESSING',
# meta={'stage':'pre_write', 'image':index+1,
# 'total':len(images)})
#write_cache(scene_cpp.cpu_cache, 1)
self.update_state(state='PROCESSING',
meta={
'stage': 'post_write',
'image': index + 1,
'total': len(images)
})
self.update_state(state='PROCESSING',
meta={'stage': 'compute error'})
normalize_3d_point_and_cov(scene_cpp.scene, scene_cpp.cpu_cache)
self.update_state(state='PROCESSING',
meta={
'stage': 'pre_write',
'image': index + 1,
'total': len(images)
})
write_cache(scene_cpp.cpu_cache, 1)
self.update_state(state='EXPORTING', meta={'stage': 'ply'})
with voxel_globe.tools.storage_dir('generate_error_point_cloud') \
as storage_dir:
ply_filename = os.path.join(storage_dir, 'model.ply')
gen_error_point_cloud(scene_cpp.scene, scene_cpp.cpu_cache,
ply_filename, prob)
potree_filename = os.path.join(storage_dir, 'potree.ply')
with voxel_globe.tools.image_dir('point_cloud') as potree_dir:
convert_ply_to_potree(ply_filename, potree_dir)
point_cloud = models.PointCloud(name='%s point cloud' %
image_set.name,
service_id=self.request.id,
origin=voxel_world.origin)
point_cloud.filename_path = ply_filename
point_cloud.potree_dir = potree_dir
point_cloud.save()
voxel_files = lambda x: glob(os.path.join(voxel_world_dir, x))
cleanup_files = []
cleanup_files += voxel_files('boxm2_covariance_*.bin')
cleanup_files += voxel_files('boxm2_point_*.bin')
cleanup_files += voxel_files('float16_image_*.bin')
for cleanup_file in cleanup_files:
os.remove(cleanup_file)
def run_build_voxel_model(self, image_collection_id, scene_id, bbox,
skip_frames, cleanup=True, history=None):
from distutils.dir_util import remove_tree
from shutil import move
import random
from vsi.tools.redirect import Redirect, Logger as LoggerWrapper
from voxel_globe.meta import models
from voxel_globe.tools.camera import get_krt
import voxel_globe.tools
from boxm2_scene_adaptor import boxm2_scene_adaptor
from vil_adaptor import load_image
from vpgl_adaptor import load_perspective_camera
from voxel_globe.tools.wget import download as wget
from vsi.vxl.create_scene_xml import create_scene_xml
from vsi.tools.dir_util import copytree, mkdtemp
with Redirect(stdout_c=LoggerWrapper(logger, lvl=logging.INFO),
stderr_c=LoggerWrapper(logger, lvl=logging.WARNING)):
openclDevice = os.environ['VIP_OPENCL_DEVICE']
opencl_memory = os.environ.get('VIP_OPENCL_MEMORY', None)
scene = models.Scene.objects.get(id=scene_id)
imageCollection = models.ImageCollection.objects.get(\
id=image_collection_id).history(history);
imageList = imageCollection.images.all();
with voxel_globe.tools.task_dir('voxel_world') as processing_dir:
logger.warning(bbox)
if bbox['geolocated']:
create_scene_xml(openclDevice, 3, float(bbox['voxel_size']),
lla1=(float(bbox['x_min']), float(bbox['y_min']),
float(bbox['z_min'])),
lla2=(float(bbox['x_max']), float(bbox['y_max']),
float(bbox['z_max'])),
origin=scene.origin, model_dir='.', number_bins=1,
output_file=open(os.path.join(processing_dir, 'scene.xml'), 'w'),
n_bytes_gpu=opencl_memory)
else:
create_scene_xml(openclDevice, 3, float(bbox['voxel_size']),
lvcs1=(float(bbox['x_min']), float(bbox['y_min']),
float(bbox['z_min'])),
lvcs2=(float(bbox['x_max']), float(bbox['y_max']),
float(bbox['z_max'])),
origin=scene.origin, model_dir='.', number_bins=1,
output_file=open(os.path.join(processing_dir, 'scene.xml'), 'w'),
n_bytes_gpu=opencl_memory)
counter = 1;
imageNames = []
cameraNames = []
os.mkdir(os.path.join(processing_dir, 'local'))
#Prepping
for image in imageList:
self.update_state(state='INITIALIZE', meta={'stage':'image fetch',
'i':counter,
'total':len(imageList)})
image = image.history(history)
(K,R,T,o) = get_krt(image.history(history), history=history)
krtName = os.path.join(processing_dir, 'local', 'frame_%05d.krt' % counter)
with open(krtName, 'w') as fid:
print >>fid, (("%0.18f "*3+"\n")*3) % (K[0,0], K[0,1], K[0,2],
K[1,0], K[1,1], K[1,2], K[2,0], K[2,1], K[2,2]);
print >>fid, (("%0.18f "*3+"\n")*3) % (R[0,0], R[0,1], R[0,2],
R[1,0], R[1,1], R[1,2], R[2,0], R[2,1], R[2,2]);
print >>fid, ("%0.18f "*3+"\n") % (T[0,0], T[1,0], T[2,0]);
imageName = image.originalImageUrl;
extension = os.path.splitext(imageName)[1]
localName = os.path.join(processing_dir, 'local',
'frame_%05d%s' % (counter, extension));
wget(imageName, localName, secret=True)
counter += 1;
imageNames.append(localName)
cameraNames.append(krtName)
variance = 0.06
vxl_scene = boxm2_scene_adaptor(os.path.join(processing_dir, "scene.xml"),
openclDevice);
current_level = 0;
loaded_imgs = [];
loaded_cams = [];
for i in range(0, len(imageNames), skip_frames):
logger.debug("i: %d img name: %s cam name: %s", i, imageNames[i],
cameraNames[i])
self.update_state(state='PRELOADING', meta={'stage':'image load',
'i':i,
'total':len(imageNames)})
img, ni, nj = load_image(imageNames[i])
loaded_imgs.append(img)
pcam = load_perspective_camera(cameraNames[i])
loaded_cams.append(pcam)
refine_cnt = 5;
for rfk in range(0, refine_cnt, 1):
pair = zip(loaded_imgs, loaded_cams)
random.shuffle(pair)
for idx, (img, cam) in enumerate(pair):
self.update_state(state='PROCESSING', meta={'stage':'update',
'i':rfk+1, 'total':refine_cnt, 'image':idx+1,
'images':len(loaded_imgs)})
logger.debug("refine_cnt: %d, idx: %d", rfk, idx)
vxl_scene.update(cam,img,True,True,None,openclDevice[0:3],variance,
tnear = 1000.0, tfar = 100000.0);
logger.debug("writing cache: %d", rfk)
vxl_scene.write_cache();
logger.debug("wrote cache: %d", rfk)
if rfk < refine_cnt-1:
self.update_state(state='PROCESSING', meta={'stage':'refine',
'i':rfk,
'total':refine_cnt})
logger.debug("refining %d...", rfk)
refine_device = openclDevice[0:3]
if refine_device == 'cpu':
refine_device = 'cpp'
vxl_scene.refine(0.3, refine_device);
vxl_scene.write_cache();
voxel_world_dir = mkdtemp(dir=os.environ['VIP_STORAGE_DIR'])
copytree(processing_dir, voxel_world_dir, ignore=lambda x,y:['images'])
models.VoxelWorld.create(
name='%s world (%s)' % (imageCollection.name, self.request.id),
origin=scene.origin,
directory=voxel_world_dir,
service_id=self.request.id).save();
def runVisualSfm(self, imageCollectionId, sceneId, cleanup=True, history=None):
from voxel_globe.meta import models
from voxel_globe.order.visualsfm.models import Order
from os import environ as env
from os.path import join as path_join
import os
import shutil
from .tools import writeNvm, writeGcpFile, generateMatchPoints, runSparse,\
readNvm
import voxel_globe.tools
from voxel_globe.tools.wget import download as wget
from voxel_globe.tools.camera import get_kto
import voxel_globe.tools.enu as enu
import numpy
import boxm2_adaptor
import boxm2_scene_adaptor
from voxel_globe.tools.xml_dict import load_xml
from django.contrib.gis.geos import Point
from voxel_globe.tools.image import convert_image
from distutils.spawn import find_executable
from glob import glob
self.update_state(state='INITIALIZE', meta={'stage':0})
#Make main temp dir and cd into it
with voxel_globe.tools.task_dir('visualsfm', cd=True) as processing_dir:
#Because visualsfm is so... bad, I have to copy it locally so I can
#configure it
visualsfm_exe = os.path.join(processing_dir,
os.path.basename(os.environ['VIP_VISUALSFM_EXE']))
shutil.copy(find_executable(os.environ['VIP_VISUALSFM_EXE']),
visualsfm_exe)
with open(os.path.join(processing_dir, 'nv.ini'), 'w') as fid:
fid.write('param_search_multiple_models 0\n')
fid.write('param_use_siftgpu 2\n')
matchFilename = path_join(processing_dir, 'match.nvm');
sparce_filename = path_join(processing_dir, 'sparse.nvm');
#This can NOT be changed in version 0.5.25
gcpFilename = matchFilename + '.gcp'
logger.debug('Task %s is processing in %s' % (self.request.id,
processing_dir))
image_collection = models.ImageCollection.objects.get(
id=imageCollectionId).history(history);
imageList = image_collection.images.all();
#A Little bit of database logging
oid = Order(processingDir=processing_dir, imageCollection=image_collection)
### if 1:
### try: #Not fully integrated yet
### sift_gpu = siftgpu.SiftGPU()
### except:
### pass
localImageList = [];
for x in range(len(imageList)):
#Download the image locally
image = imageList[x].history(history);
self.update_state(state='INITIALIZE', meta={'stage':'image fetch', 'i':x,
'total':len(imageList)})
imageName = image.originalImageUrl;
extension = os.path.splitext(imageName)[1].lower()
localName = path_join(processing_dir, 'frame_%05d%s' % (x+1, extension));
wget(imageName, localName, secret=True)
#Convert the image if necessary
if extension not in ['.jpg', '.jpeg', '.pgm', '.ppm']:
self.update_state(state='INITIALIZE',
meta={'stage':'image convert', 'i':x, 'total':len(imageList)})
#Add code here to converty to jpg for visual sfm
if extension in ['.png']:#'not implemented':
from PIL import Image
image_temp = Image.open(localName)
if len(image_temp.mode) > 1: #Stupid visual sfm is picky :(
new_local_name = os.path.splitext(localName)[0] + '.ppm';
else:
new_local_name = os.path.splitext(localName)[0] + '.pgm';
new_local_name = os.path.splitext(localName)[0] + '.jpg';
###ingest.convert_image(localName, new_local_name, 'PNM')
convert_image(localName, new_local_name, 'JPEG',
options=('QUALITY=100',))
os.remove(localName)
localName = new_local_name;
else:
raise Exception('Unsupported file type');
imageInfo = {'localName':localName, 'index':x}
try:
[K, T, llh] = get_kto(image, history=history);
imageInfo['K_intrinsics'] = K;
imageInfo['transformation'] = T;
imageInfo['enu_origin'] = llh;
except:
pass
localImageList.append(imageInfo);
### if 1:
### try: #not fully integrated yet
### sift_gpu.create_sift(localName, os.path.splitext(localName)[0]+'.sift')
### except:
### pass
# filenames = list(imageList.values_list('imageUrl'))
# logger.info('The image list 0is %s' % filenames)
self.update_state(state='PROCESSING',
meta={'stage':'generate match points',
'processing_dir':processing_dir,
'total':len(imageList)})
generateMatchPoints(map(lambda x:x['localName'], localImageList),
matchFilename, logger=logger, executable=visualsfm_exe)
# cameras = [];
# for image in imageList:
# if 1:
# #try:
# [K, T, llh] = get_kto(image);
# cameras.append({'image':image.id, 'K':K, 'tranformation':
# T, 'origin':llh})
# #except:
# pass
# origin = numpy.median(origin, axis=0)
# origin = [-92.215197, 37.648858, 268.599]
scene = models.Scene.objects.get(id=sceneId).history(history)
origin = list(scene.origin)
if scene.geolocated:
self.update_state(state='PROCESSING',
meta={'stage':'writing gcp points'})
#find the middle origin, and make it THE origin
data = []#.name .llh_xyz
for imageInfo in localImageList:
try:
r = imageInfo['transformation'][0:3, 0:3]
t = imageInfo['transformation'][0:3, 3:]
enu_point = -r.transpose().dot(t);
if not numpy.array_equal(imageInfo['enu_origin'], origin):
ecef = enu.enu2xyz(refLong=imageInfo['enu_origin'][0],
refLat=imageInfo['enu_origin'][1],
refH=imageInfo['enu_origin'][2],
#e=imageInfo['transformation'][0, 3],
#n=imageInfo['transformation'][1, 3],
#u=imageInfo['transformation'][2, 3])
e=enu_point[0],
n=enu_point[1],
u=enu_point[2])
enu_point = enu.xyz2enu(refLong=origin[0],
refLat=origin[1],
refH=origin[2],
X=ecef[0],
Y=ecef[1],
Z=ecef[2])
# else:
# enu_point = imageInfo['transformation'][0:3, 3];
dataBit = {'filename':imageInfo['localName'], 'xyz':enu_point}
data.append(dataBit);
#Make this a separate ingest process, making CAMERAS linked to the
#images
#data = arducopter.loadAdjTaggedMetadata(
# r'd:\visualsfm\2014-03-20 13-22-44_adj_tagged_images.txt');
#Make this read the cameras from the DB instead
writeGcpFile(data, gcpFilename)
except: #some images may have no camera
pass
oid.lvcsOrigin = str(origin)
oid.save()
self.update_state(state='PROCESSING', meta={'stage':'sparse SFM'})
runSparse(matchFilename, sparce_filename, gcp=scene.geolocated,
shared=True, logger=logger, executable=visualsfm_exe)
self.update_state(state='FINALIZE',
meta={'stage':'loading resulting cameras'})
#prevent bundle2scene from getting confused and crashing
sift_data = os.path.join(processing_dir, 'sift_data')
os.mkdir(sift_data)
for filename in glob(os.path.join(processing_dir, '*.mat')) +\
glob(os.path.join(processing_dir, '*.sift')):
shutil.move(filename, sift_data)
if scene.geolocated:
#Create a uscene.xml for the geolocated case. All I want out of this is
#the bounding box and gsd calculation.
boxm2_adaptor.bundle2scene(sparce_filename, processing_dir, isalign=False,
out_dir="")
cams = readNvm(path_join(processing_dir, 'sparse.nvm'))
#cams.sort(key=lambda x:x.name)
#Since the file names are frame_00001, etc... and you KNOW this order is
#identical to localImageList, with some missing
for cam in cams:
frameName = cam.name; #frame_00001, etc....
imageInfo = filter(lambda x: x['localName'].endswith(frameName),
localImageList)[0]
#I have to use endswith instead of == because visual sfm APPARENTLY
#decides to take some liberty and make absolute paths relative
image = imageList[imageInfo['index']].history(history)
(k,r,t) = cam.krt(width=image.imageWidth, height=image.imageHeight);
logger.info('Origin is %s' % str(origin))
llh_xyz = enu.enu2llh(lon_origin=origin[0],
lat_origin=origin[1],
h_origin=origin[2],
east=cam.translation_xyz[0],
north=cam.translation_xyz[1],
up=cam.translation_xyz[2])
grcs = models.GeoreferenceCoordinateSystem.create(
name='%s 0' % image.name,
xUnit='d', yUnit='d', zUnit='m',
location='SRID=4326;POINT(%0.15f %0.15f %0.15f)'
% (origin[0], origin[1], origin[2]),
service_id = self.request.id)
grcs.save()
cs = models.CartesianCoordinateSystem.create(
name='%s 1' % (image.name),
service_id = self.request.id,
xUnit='m', yUnit='m', zUnit='m');
cs.save()
transform = models.CartesianTransform.create(
name='%s 1_0' % (image.name),
service_id = self.request.id,
rodriguezX=Point(*r[0,:]),
rodriguezY=Point(*r[1,:]),
rodriguezZ=Point(*r[2,:]),
translation=Point(t[0][0], t[1][0], t[2][0]),
coordinateSystem_from_id=grcs.id,
coordinateSystem_to_id=cs.id)
transform.save()
camera = image.camera;
try:
camera.update(service_id = self.request.id,
focalLengthU=k[0,0], focalLengthV=k[1,1],
principalPointU=k[0,2], principalPointV=k[1,2],
coordinateSystem=cs);
except:
camera = models.Camera.create(name=image.name,
service_id = self.request.id,
focalLengthU=k[0,0], focalLengthV=k[1,1],
principalPointU=k[0,2], principalPointV=k[1,2],
coordinateSystem=cs);
camera.save();
image.update(camera = camera);
logger.info(str(cams[0]))
else:
from vsi.tools.natural_sort import natural_sorted
from glob import glob
from vsi.io.krt import Krt
from voxel_globe.tools.camera import save_krt
boxm2_adaptor.bundle2scene(sparce_filename, processing_dir, isalign=True,
out_dir=processing_dir)
#While the output dir is used for the b2s folders, uscene.xml is cwd
#They are both set to processing_dir, so everything works out well
aligned_cams = glob(os.path.join(processing_dir, 'cams_krt', '*'))
#sort them naturally in case there are more then 99,999 files
aligned_cams = natural_sorted(aligned_cams)
if len(aligned_cams) != len(imageList):
#Create a new image collection
new_image_collection = models.ImageCollection.create(
name="SFM Result Subset (%s)" % image_collection.name,
service_id = self.request.id);
# for image in image_collection.images.all():
# new_image_collection.images.add(image)
new_image_collection.save();
frames_keep = set(map(lambda x:
int(os.path.splitext(x.split('_')[-2])[0])-1, aligned_cams))
for frame_index in frames_keep:
new_image_collection.images.add(imageList[frame_index])
# frames_remove = set(xrange(len(imageList))) - frames_keep
#
# for remove_index in list(frames_remove):
# #The frame number refers to the nth image in the image collection,
# #so frame_00100.tif is the 100th image, starting the index at one
# #See local_name above
#
# #remove the images sfm threw away
# new_image_collection.remove(imageList[remove_index])
image_collection = new_image_collection
frames_keep = list(frames_keep)
else:
frames_keep = xrange(len(aligned_cams))
#---Update the camera models in the database.---
for camera_index, frame_index in enumerate(frames_keep):
krt = Krt.load(aligned_cams[camera_index])
image = imageList[frame_index].history(history)
save_krt(self.request.id, image, krt.k, krt.r, krt.t, [0,0,0],
srid=4326)
#---Update scene information important for the no-metadata case ---
scene_filename = os.path.join(processing_dir, 'model', 'uscene.xml')
boxm_scene = boxm2_scene_adaptor.boxm2_scene_adaptor(scene_filename)
scene.bbox_min = 'POINT(%0.15f %0.15f %0.15f)' % boxm_scene.bbox[0]
scene.bbox_max = 'POINT(%0.15f %0.15f %0.15f)' % boxm_scene.bbox[1]
#This is not a complete or good function really... but it will get me the
#information I need.
scene_dict = load_xml(scene_filename)
block = scene_dict['block']
scene.default_voxel_size='POINT(%f %f %f)' % \
(float(block.at['dim_x']), float(block.at['dim_y']),
float(block.at['dim_z']))
scene.save()
return oid.id;
def run_build_voxel_model(self, image_set_id, camera_set_id, scene_id, bbox,
skip_frames, cleanup=True):
import random
from vsi.tools.redirect import StdRedirect
from voxel_globe.meta import models
from voxel_globe.tools.camera import get_krt
import voxel_globe.tools
from boxm2_scene_adaptor import boxm2_scene_adaptor
import brl_init
from vil_adaptor_boxm2_batch import load_image
from vpgl_adaptor_boxm2_batch import load_perspective_camera
from vsi.vxl.create_scene_xml import create_scene_xml
from vsi.tools.dir_util import copytree
from vsi.tools.file_util import lncp
with StdRedirect(open(os.path.join(voxel_globe.tools.log_dir(),
self.request.id)+'_out.log', 'w'),
open(os.path.join(voxel_globe.tools.log_dir(),
self.request.id)+'_err.log', 'w')):
openclDevice = os.environ['VIP_OPENCL_DEVICE']
opencl_memory = os.environ.get('VIP_OPENCL_MEMORY', None)
if opencl_memory:
opencl_memory = int(opencl_memory)
scene = models.Scene.objects.get(id=scene_id)
imageSet = models.ImageSet.objects.get(\
id=image_set_id)
imageList = imageSet.images.all()
with voxel_globe.tools.task_dir('voxel_world') as processing_dir:
logger.warning(bbox)
create_scene_xml(openclDevice, 3, float(bbox['voxel_size']),
lvcs1=(float(bbox['x_min']), float(bbox['y_min']),
float(bbox['z_min'])),
lvcs2=(float(bbox['x_max']), float(bbox['y_max']),
float(bbox['z_max'])),
origin=scene.origin, model_dir='.', number_bins=1,
output_file=open(os.path.join(processing_dir, 'scene.xml'), 'w'),
n_bytes_gpu=opencl_memory)
counter = 1
imageNames = []
cameraNames = []
os.mkdir(os.path.join(processing_dir, 'local'))
#Prepping
for image in imageList:
self.update_state(state='INITIALIZE', meta={'image_set_name': imageSet.name,
'stage':'image fetch',
'i':counter,
'total':len(imageList)})
(K,R,T,o) = get_krt(image, camera_set_id)
krtName = os.path.join(processing_dir, 'local', 'frame_%05d.krt' % counter)
with open(krtName, 'w') as fid:
print >>fid, (("%0.18f "*3+"\n")*3) % (K[0,0], K[0,1], K[0,2],
K[1,0], K[1,1], K[1,2], K[2,0], K[2,1], K[2,2])
print >>fid, (("%0.18f "*3+"\n")*3) % (R[0,0], R[0,1], R[0,2],
R[1,0], R[1,1], R[1,2], R[2,0], R[2,1], R[2,2])
print >>fid, ("%0.18f "*3+"\n") % (T[0,0], T[1,0], T[2,0])
imageName = image.filename_path
extension = os.path.splitext(imageName)[1]
localName = os.path.join(processing_dir, 'local',
'frame_%05d%s' % (counter, extension))
lncp(imageName, localName)
counter += 1
imageNames.append(localName)
cameraNames.append(krtName)
variance = 0.06
vxl_scene = boxm2_scene_adaptor(os.path.join(processing_dir, "scene.xml"),
openclDevice)
loaded_imgs = []
loaded_cams = []
for i in range(0, len(imageNames), skip_frames):
logger.debug("i: %d img name: %s cam name: %s", i, imageNames[i],
cameraNames[i])
self.update_state(state='PRELOADING', meta={'image_set_name': imageSet.name,
'stage':'image load',
'i':i,
'total':len(imageNames)})
img, ni, nj = load_image(imageNames[i])
loaded_imgs.append(img)
pcam = load_perspective_camera(cameraNames[i])
loaded_cams.append(pcam)
refine_cnt = 5
for rfk in range(0, refine_cnt, 1):
pair = zip(loaded_imgs, loaded_cams)
random.shuffle(pair)
for idx, (img, cam) in enumerate(pair):
self.update_state(state='PROCESSING', meta={'image_set_name': imageSet.name,
'stage':'update',
'i':rfk+1, 'total':refine_cnt, 'image':idx+1,
'images':len(loaded_imgs)})
logger.debug("refine_cnt: %d, idx: %d", rfk, idx)
vxl_scene.update(cam,img,True,True,None,openclDevice,variance,
tnear = 1000.0, tfar = 100000.0)
logger.debug("writing cache: %d", rfk)
vxl_scene.write_cache()
logger.debug("wrote cache: %d", rfk)
if rfk < refine_cnt-1:
self.update_state(state='PROCESSING', meta={'image_set_name': imageSet.name,
'stage':'refine',
'i':rfk,
'total':refine_cnt})
logger.debug("refining %d...", rfk)
vxl_scene.refine(0.3, openclDevice)
vxl_scene.write_cache()
with open(os.path.join(processing_dir, "scene_color.xml"), 'w') as fid:
lines = open(os.path.join(processing_dir, "scene.xml"),
'r').readlines()
lines = [line.replace('boxm2_mog3_grey',
'boxm2_gauss_rgb').replace(
'boxm2_num_obs',
'boxm2_num_obs_single') for line in lines]
fid.writelines(lines)
vxl_scene = boxm2_scene_adaptor(os.path.join(processing_dir,
"scene_color.xml"),
openclDevice)
for idx, (img, cam) in enumerate(pair):
self.update_state(state='PROCESSING', meta={'image_set_name': imageSet.name,
'stage':'color_update',
'i':rfk+1, 'total':refine_cnt, 'image':idx+1,
'images':len(loaded_imgs)})
logger.debug("color_paint idx: %d", idx)
vxl_scene.update(cam,img,False,False,None,openclDevice,
tnear = 1000.0, tfar = 100000.0)
vxl_scene.write_cache()
with voxel_globe.tools.storage_dir('voxel_world') as voxel_world_dir:
copytree(processing_dir, voxel_world_dir, ignore=lambda x,y:['local'])
models.VoxelWorld(
name='%s world (%s)' % (imageSet.name, self.request.id),
origin=scene.origin,
directory=voxel_world_dir,
service_id=self.request.id).save()
return {"image_set_name" : imageSet.name}
def runVisualSfm(self, imageCollectionId, sceneId, cleanup=True, history=None):
from voxel_globe.meta import models
from voxel_globe.order.visualsfm.models import Order
from os import environ as env
from os.path import join as path_join
import os
import shutil
from .tools import writeNvm, writeGcpFile, generateMatchPoints, runSparse,\
readNvm
import voxel_globe.tools
from voxel_globe.tools.wget import download as wget
from voxel_globe.tools.camera import get_kto
import voxel_globe.tools.enu as enu
import numpy
import boxm2_adaptor
import boxm2_scene_adaptor
from voxel_globe.tools.xml_dict import load_xml
from django.contrib.gis.geos import Point
from voxel_globe.tools.image import convert_image
from distutils.spawn import find_executable
from glob import glob
self.update_state(state='INITIALIZE', meta={'stage': 0})
#Make main temp dir and cd into it
with voxel_globe.tools.task_dir('visualsfm', cd=True) as processing_dir:
#Because visualsfm is so... bad, I have to copy it locally so I can
#configure it
visualsfm_exe = os.path.join(
processing_dir, os.path.basename(os.environ['VIP_VISUALSFM_EXE']))
shutil.copy(find_executable(os.environ['VIP_VISUALSFM_EXE']),
visualsfm_exe)
with open(os.path.join(processing_dir, 'nv.ini'), 'w') as fid:
fid.write('param_search_multiple_models 0\n')
fid.write('param_use_siftgpu 2\n')
matchFilename = path_join(processing_dir, 'match.nvm')
sparce_filename = path_join(processing_dir, 'sparse.nvm')
#This can NOT be changed in version 0.5.25
gcpFilename = matchFilename + '.gcp'
logger.debug('Task %s is processing in %s' %
(self.request.id, processing_dir))
image_collection = models.ImageCollection.objects.get(
id=imageCollectionId).history(history)
imageList = image_collection.images.all()
#A Little bit of database logging
oid = Order(processingDir=processing_dir,
imageCollection=image_collection)
### if 1:
### try: #Not fully integrated yet
### sift_gpu = siftgpu.SiftGPU()
### except:
### pass
localImageList = []
for x in range(len(imageList)):
#Download the image locally
image = imageList[x].history(history)
self.update_state(state='INITIALIZE',
meta={
'stage': 'image fetch',
'i': x,
'total': len(imageList)
})
imageName = image.originalImageUrl
extension = os.path.splitext(imageName)[1].lower()
localName = path_join(processing_dir,
'frame_%05d%s' % (x + 1, extension))
wget(imageName, localName, secret=True)
#Convert the image if necessary
if extension not in ['.jpg', '.jpeg', '.pgm', '.ppm']:
self.update_state(state='INITIALIZE',
meta={
'stage': 'image convert',
'i': x,
'total': len(imageList)
})
#Add code here to converty to jpg for visual sfm
if extension in ['.png']: #'not implemented':
from PIL import Image
image_temp = Image.open(localName)
if len(image_temp.mode
) > 1: #Stupid visual sfm is picky :(
new_local_name = os.path.splitext(
localName)[0] + '.ppm'
else:
new_local_name = os.path.splitext(
localName)[0] + '.pgm'
new_local_name = os.path.splitext(localName)[0] + '.jpg'
###ingest.convert_image(localName, new_local_name, 'PNM')
convert_image(localName,
new_local_name,
'JPEG',
options=('QUALITY=100', ))
os.remove(localName)
localName = new_local_name
else:
raise Exception('Unsupported file type')
imageInfo = {'localName': localName, 'index': x}
try:
[K, T, llh] = get_kto(image, history=history)
imageInfo['K_intrinsics'] = K
imageInfo['transformation'] = T
imageInfo['enu_origin'] = llh
except:
pass
localImageList.append(imageInfo)
### if 1:
### try: #not fully integrated yet
### sift_gpu.create_sift(localName, os.path.splitext(localName)[0]+'.sift')
### except:
### pass
# filenames = list(imageList.values_list('imageUrl'))
# logger.info('The image list 0is %s' % filenames)
self.update_state(state='PROCESSING',
meta={
'stage': 'generate match points',
'processing_dir': processing_dir,
'total': len(imageList)
})
generateMatchPoints(map(lambda x: x['localName'], localImageList),
matchFilename,
logger=logger,
executable=visualsfm_exe)
# cameras = [];
# for image in imageList:
# if 1:
# #try:
# [K, T, llh] = get_kto(image);
# cameras.append({'image':image.id, 'K':K, 'tranformation':
# T, 'origin':llh})
# #except:
# pass
# origin = numpy.median(origin, axis=0)
# origin = [-92.215197, 37.648858, 268.599]
scene = models.Scene.objects.get(id=sceneId).history(history)
origin = list(scene.origin)
if scene.geolocated:
self.update_state(state='PROCESSING',
meta={'stage': 'writing gcp points'})
#find the middle origin, and make it THE origin
data = [] #.name .llh_xyz
for imageInfo in localImageList:
try:
r = imageInfo['transformation'][0:3, 0:3]
t = imageInfo['transformation'][0:3, 3:]
enu_point = -r.transpose().dot(t)
if not numpy.array_equal(imageInfo['enu_origin'], origin):
ecef = enu.enu2xyz(
refLong=imageInfo['enu_origin'][0],
refLat=imageInfo['enu_origin'][1],
refH=imageInfo['enu_origin'][2],
#e=imageInfo['transformation'][0, 3],
#n=imageInfo['transformation'][1, 3],
#u=imageInfo['transformation'][2, 3])
e=enu_point[0],
n=enu_point[1],
u=enu_point[2])
enu_point = enu.xyz2enu(refLong=origin[0],
refLat=origin[1],
refH=origin[2],
X=ecef[0],
Y=ecef[1],
Z=ecef[2])
# else:
# enu_point = imageInfo['transformation'][0:3, 3];
dataBit = {
'filename': imageInfo['localName'],
'xyz': enu_point
}
data.append(dataBit)
#Make this a separate ingest process, making CAMERAS linked to the
#images
#data = arducopter.loadAdjTaggedMetadata(
# r'd:\visualsfm\2014-03-20 13-22-44_adj_tagged_images.txt');
#Make this read the cameras from the DB instead
writeGcpFile(data, gcpFilename)
except: #some images may have no camera
pass
oid.lvcsOrigin = str(origin)
oid.save()
self.update_state(state='PROCESSING', meta={'stage': 'sparse SFM'})
runSparse(matchFilename,
sparce_filename,
gcp=scene.geolocated,
shared=True,
logger=logger,
executable=visualsfm_exe)
self.update_state(state='FINALIZE',
meta={'stage': 'loading resulting cameras'})
#prevent bundle2scene from getting confused and crashing
sift_data = os.path.join(processing_dir, 'sift_data')
os.mkdir(sift_data)
for filename in glob(os.path.join(processing_dir, '*.mat')) +\
glob(os.path.join(processing_dir, '*.sift')):
shutil.move(filename, sift_data)
if scene.geolocated:
#Create a uscene.xml for the geolocated case. All I want out of this is
#the bounding box and gsd calculation.
boxm2_adaptor.bundle2scene(sparce_filename,
processing_dir,
isalign=False,
out_dir="")
cams = readNvm(path_join(processing_dir, 'sparse.nvm'))
#cams.sort(key=lambda x:x.name)
#Since the file names are frame_00001, etc... and you KNOW this order is
#identical to localImageList, with some missing
for cam in cams:
frameName = cam.name
#frame_00001, etc....
imageInfo = filter(
lambda x: x['localName'].endswith(frameName),
localImageList)[0]
#I have to use endswith instead of == because visual sfm APPARENTLY
#decides to take some liberty and make absolute paths relative
image = imageList[imageInfo['index']].history(history)
(k, r, t) = cam.krt(width=image.imageWidth,
height=image.imageHeight)
logger.info('Origin is %s' % str(origin))
llh_xyz = enu.enu2llh(lon_origin=origin[0],
lat_origin=origin[1],
h_origin=origin[2],
east=cam.translation_xyz[0],
north=cam.translation_xyz[1],
up=cam.translation_xyz[2])
grcs = models.GeoreferenceCoordinateSystem.create(
name='%s 0' % image.name,
xUnit='d',
yUnit='d',
zUnit='m',
location='SRID=4326;POINT(%0.15f %0.15f %0.15f)' %
(origin[0], origin[1], origin[2]),
service_id=self.request.id)
grcs.save()
cs = models.CartesianCoordinateSystem.create(
name='%s 1' % (image.name),
service_id=self.request.id,
xUnit='m',
yUnit='m',
zUnit='m')
cs.save()
transform = models.CartesianTransform.create(
name='%s 1_0' % (image.name),
service_id=self.request.id,
rodriguezX=Point(*r[0, :]),
rodriguezY=Point(*r[1, :]),
rodriguezZ=Point(*r[2, :]),
translation=Point(t[0][0], t[1][0], t[2][0]),
coordinateSystem_from_id=grcs.id,
coordinateSystem_to_id=cs.id)
transform.save()
camera = image.camera
try:
camera.update(service_id=self.request.id,
focalLengthU=k[0, 0],
focalLengthV=k[1, 1],
principalPointU=k[0, 2],
principalPointV=k[1, 2],
coordinateSystem=cs)
except:
camera = models.Camera.create(name=image.name,
service_id=self.request.id,
focalLengthU=k[0, 0],
focalLengthV=k[1, 1],
principalPointU=k[0, 2],
principalPointV=k[1, 2],
coordinateSystem=cs)
camera.save()
image.update(camera=camera)
logger.info(str(cams[0]))
else:
from vsi.tools.natural_sort import natural_sorted
from glob import glob
from vsi.io.krt import Krt
from voxel_globe.tools.camera import save_krt
boxm2_adaptor.bundle2scene(sparce_filename,
processing_dir,
isalign=True,
out_dir=processing_dir)
#While the output dir is used for the b2s folders, uscene.xml is cwd
#They are both set to processing_dir, so everything works out well
aligned_cams = glob(os.path.join(processing_dir, 'cams_krt', '*'))
#sort them naturally in case there are more then 99,999 files
aligned_cams = natural_sorted(aligned_cams)
if len(aligned_cams) != len(imageList):
#Create a new image collection
new_image_collection = models.ImageCollection.create(
name="SFM Result Subset (%s)" % image_collection.name,
service_id=self.request.id)
# for image in image_collection.images.all():
# new_image_collection.images.add(image)
new_image_collection.save()
frames_keep = set(
map(
lambda x: int(os.path.splitext(x.split('_')[-2])[0]) -
1, aligned_cams))
for frame_index in frames_keep:
new_image_collection.images.add(imageList[frame_index])
# frames_remove = set(xrange(len(imageList))) - frames_keep
#
# for remove_index in list(frames_remove):
# #The frame number refers to the nth image in the image collection,
# #so frame_00100.tif is the 100th image, starting the index at one
# #See local_name above
#
# #remove the images sfm threw away
# new_image_collection.remove(imageList[remove_index])
image_collection = new_image_collection
frames_keep = list(frames_keep)
else:
frames_keep = xrange(len(aligned_cams))
#---Update the camera models in the database.---
for camera_index, frame_index in enumerate(frames_keep):
krt = Krt.load(aligned_cams[camera_index])
image = imageList[frame_index].history(history)
save_krt(self.request.id,
image,
krt.k,
krt.r,
krt.t, [0, 0, 0],
srid=4326)
#---Update scene information important for the no-metadata case ---
scene_filename = os.path.join(processing_dir, 'model', 'uscene.xml')
boxm_scene = boxm2_scene_adaptor.boxm2_scene_adaptor(scene_filename)
scene.bbox_min = 'POINT(%0.15f %0.15f %0.15f)' % boxm_scene.bbox[0]
scene.bbox_max = 'POINT(%0.15f %0.15f %0.15f)' % boxm_scene.bbox[1]
#This is not a complete or good function really... but it will get me the
#information I need.
scene_dict = load_xml(scene_filename)
block = scene_dict['block']
scene.default_voxel_size='POINT(%f %f %f)' % \
(float(block.at['dim_x']), float(block.at['dim_y']),
float(block.at['dim_z']))
scene.save()
return oid.id
def create_height_map(self, voxel_world_id, render_height):
import shutil
import urllib
import numpy as np
import brl_init
from boxm2_scene_adaptor import boxm2_scene_adaptor
from boxm2_adaptor import ortho_geo_cam_from_scene, scene_lvcs, scene_bbox
from vpgl_adaptor_boxm2_batch import convert_local_to_global_coordinates, geo2generic, save_geocam_to_tfw
from vil_adaptor_boxm2_batch import save_image, scale_and_offset_values, stretch_image, image_range
import vsi.io.image
import voxel_globe.tools
import voxel_globe.tools.hash
import voxel_globe.tools.camera
import voxel_globe.meta.models as models
import voxel_globe.ingest.payload.tools
with voxel_globe.tools.task_dir('height_map', cd=True) as processing_dir:
voxel_world = models.VoxelWorld.objects.get(id=voxel_world_id)
scene = boxm2_scene_adaptor(
os.path.join(voxel_world.directory, 'scene.xml'),
env['VIP_OPENCL_DEVICE'])
ortho_camera, cols, rows = ortho_geo_cam_from_scene(scene.scene)
tfw_camera = os.path.join(processing_dir, 'cam.tfw')
save_geocam_to_tfw(ortho_camera, tfw_camera)
with open(tfw_camera, 'r') as fid:
geo_transform = [float(x) for x in fid.readlines()]
(x0, y0, z0), (x1, y1, z1) = scene_bbox(scene.scene)
lvcs = scene_lvcs(scene.scene)
#lvcs = vpgl_adaptor.create_lvcs(lat=origin[1], lon=origin[0], el=origin[2],
# csname="wgs84")
_, _, min_height = convert_local_to_global_coordinates(
lvcs, x0, y0, z0)
if render_height is None:
render_height = z1 + (z1 - z0) / 1000
#z1+(z1-z0)/1000 is basically to say "just a little above the top" *2 is
#1) overkill and 2) doesn't work with sign, +1 could go crazy in an
#arbitrarily scaled system, so this just calculates ".1% more" which is
#more than good enough
else:
render_height = render_height - voxel_world.origin[2]
logger.critical("Render Height is %f (%s)", render_height,
type(render_height))
generic_camera = geo2generic(ortho_camera, cols, rows, render_height,
0)
z_exp_img, z_var_img = scene.render_z_image(generic_camera, cols, rows)
#This is TECHNICALLY wrong, it assumes the earth is flat.
scale_and_offset_values(z_exp_img, 1, min_height)
height_filename = os.path.join(processing_dir, 'height.tif')
save_image(z_exp_img, height_filename)
checksum = voxel_globe.tools.hash.sha256_file(height_filename)
with voxel_globe.tools.image_sha_dir(checksum) as image_dir:
original_filename = os.path.join(image_dir, 'height_map.tif')
#If the exact file exist already, don't ingest it again. Unlikely
if not os.path.exists(original_filename):
img = vsi.io.image.imread(height_filename)
vsi.io.image.imwrite_geotiff(
img.raster(), original_filename,
[geo_transform[x] for x in [4, 0, 1, 5, 2, 3]])
zoomify_filename = os.path.join(image_dir, 'zoomify.tif')
img_min, img_max = image_range(z_exp_img)
if img_min == img_max:
zoomify_image = z_exp_img #At least it won't crash
else:
zoomify_image = stretch_image(z_exp_img, img_min, img_max,
'byte')
save_image(zoomify_image, zoomify_filename)
zoomify_name = os.path.join(image_dir, 'zoomify')
voxel_globe.ingest.payload.tools.zoomify_image(
zoomify_filename, zoomify_name)
img = voxel_globe.meta.models.Image(
name="Height Map %s (%s)" % (voxel_world.name, voxel_world.id),
image_width=cols,
image_height=rows,
number_bands=1,
pixel_format='f',
file_format='zoom',
service_id=self.request.id)
img.filename_path = original_filename
img.save()
image_set = models.ImageSet.objects.get_or_create(
name="Height Maps",
defaults={"_attributes": '{"autogen":true}'})[0]
image_set.images.add(img)
gsd = scene.description['voxelLength']
camera_center = ((x0 + x1) / 2, (y0 + y1) / 2, z1 + 10000)
d = z1 - z0 + 10000
k = np.eye(3)
k[0, 2] = cols / 2
k[1, 2] = rows / 2
k[0, 0] = k[1, 1] = d / gsd
r = np.eye(3)
r[0, 0] = -1
t = -r.T.dot(camera_center)
camera = voxel_globe.tools.camera.save_krt(self.request.id, img, k,
r, t,
voxel_world.origin)
camera_set=voxel_globe.meta.models.CameraSet(\
name="Height Map %s (%s)" % (voxel_world.name, voxel_world.id), \
images=image_set, service_id=self.request.id)
camera_set.save()
camera_set.cameras.add(camera)
