def run(colmap_folder, sfs_results_path, gt_path, out_folder):
#import pdb;pdb.set_trace();
if not sfs_results_path.endswith('/'):
sfs_results_path += '/'
if not gt_path.endswith('/'):
gt_path += '/'
if not out_folder.endswith('/'):
out_folder += '/'
scene_manager = SceneManager(colmap_folder)
scene_manager.load_cameras()
camera = scene_manager.cameras[1] # assume single camera
# initial values on unit sphere
x, y = camera.get_image_grid()
r_scale = np.sqrt(x * x + y * y + 1.)
image = np.empty((camera.height, camera.width, 3))
#image[:,:,0] = 1.
if not os.path.exists(out_folder):
os.mkdir(out_folder)
# iterate through rendered BRDFs
#for brdf_name in os.listdir(sfs_results_path):
for brdf_name in ['phantom']:
brdf_folder = sfs_results_path + brdf_name + '/'
if not os.path.isdir(brdf_folder): continue
with open(out_folder + brdf_name + '.csv', 'w') as f:
print >> f, HEADER1 % brdf_name
print >> f, HEADER2
# iterate through reflectance models
#for rm_name in os.listdir(brdf_folder):
#for model_name in MODELS:
model_folder = brdf_folder + '/' #model_name + '/'
# percent of pixels within 1, 2, 5mm of the GT surface
data = [list() for _ in THRESHOLDS]
###
datasum = []
###
# iterate through SFS surfaces
for filename in os.listdir(model_folder):
if not filename.endswith('_z.bin'): continue
name = filename[:-6]
#import pdb;pdb.set_trace();
util.save_sfs_ply(
'%s_%s.ply' % (name, 'est'),
np.dstack((x, y, np.ones_like(x))) * np.fromfile(
model_folder + filename, dtype=np.float32).reshape(
camera.height, camera.width, 1))
#
util.save_sfs_ply(
'%s_%s.ply' % (name, 'gt'),
np.dstack(
(x, y, np.ones_like(x))) / WORLD_SCALE * np.fromfile(
gt_path + name + '.bin', dtype=np.float32).reshape(
camera.height, camera.width, 1))
#break
r_est = WORLD_SCALE * r_scale * np.fromfile(
model_folder + filename, dtype=np.float32).reshape(
camera.height, camera.width)
r_gt = r_scale * np.fromfile(gt_path + name + '.bin',
dtype=np.float32).reshape(
camera.height, camera.width)
rdiff = np.abs(r_est - r_gt)
#rdiff = np.concatenate((
# rdiff[:,:20].ravel(), rdiff[:,-20:].ravel(),
# rdiff[20:-20,:20].ravel(), rdiff[20:-20,-20:].ravel()))
#rdiff = rdiff[(r_gt > 5.0) & (r_gt < 20.0)]
inv_size = 1. / float(rdiff.size)
for i, t in enumerate(THRESHOLDS):
data[i].append(np.count_nonzero(rdiff < t) * inv_size)
#data[i].append(np.mean((rdiff / r_gt)[rdiff < t]))
#data[i].append(np.mean((r_est / r_gt)))
###
datasum.append(1. - (r_est / r_gt).ravel())
datasum = np.concatenate(datasum)
print np.mean(datasum), np.std(datasum)
###
data = np.array(data)
data = np.mean(data, axis=1), np.std(data, axis=1)
model_name = 'power'
print >> f, model_name + ',' + ','.join('%f,%f' % d
for d in izip(*data))
def run_iterative_sfs(out_file, model_type, fit_falloff, *extra_params):
#if os.path.exists(out_file + '_z.bin'): # don't re-run existing results
# return
if model_type == sfs.LAMBERTIAN_MODEL:
model_name = 'LAMBERTIAN_MODEL'
elif model_type == sfs.OREN_NAYAR_MODEL:
model_name = 'OREN_NAYAR_MODEL'
elif model_type == sfs.PHONG_MODEL:
model_name = 'PHONG_MODEL'
elif model_type == sfs.COOK_TORRANCE_MODEL:
model_name = 'COOK_TORRANCE_MODEL'
elif model_type == sfs.POWER_MODEL:
model_name = 'POWER_MODEL'
if model_type == sfs.POWER_MODEL:
print '%s_%i' % (model_name, extra_params[0])
else:
print model_name
S = S0.copy()
z = S0[:, :, 2]
for iteration in xrange(MAX_NUM_ITER):
print 'Iteration', iteration
z_old = z
if get_initial_warp:
#z_gt=util.load_point_ply('C:\\Users\\user\\Documents\\UNC\\Research\\ColonProject\\code\\Rui\\SFS_CPU\\frame0859.jpg_gt.ply')
# warp to 3D points
if iteration > -1:
S, r_ratios = nearest_neighbor_warp(
weights, nn_idxs, points2D_image, r_fixed,
util.generate_surface(camera, z))
z_est = np.maximum(S[:, :, 2], 1e-6)
S = util.generate_surface(camera, z_est)
else:
z_est = z
S = util.generate_surface(camera, z_est)
#util.save_sfs_ply('warp' + '.ply', S, im_rgb)
#util.save_xyz('test.xyz',points3D);
#z=z_est
#break
#z_est=z
else:
#import pdb;pdb.set_trace()
#z_est = extract_depth_map(camera,ref_surf_name,R,image)
z_est = np.fromfile(
'C:\Users\user\Documents\UNC\Research\ColonProject\code\SFS_Program_from_True\endo_evaluation\gt_surfaces\\frame0859.jpg.bin',
dtype=np.float32).reshape(camera.height,
camera.width) / WORLD_SCALE
z_est = z_est.astype(float)
#S, r_ratios = nearest_neighbor_warp(weights, nn_idxs,
# points2D_image, r_fixed, util.generate_surface(camera, z_est))
z_est = np.maximum(z_est[:, :], 1e-6)
#Sworld = (S - image.tvec[np.newaxis,np.newaxis,:]).dot(R)
S = util.generate_surface(camera, z_est)
#util.save_sfs_ply('test' + '.ply', S, im_rgb)
#util.save_sfs_ply(out_file + '_warp_%i.ply' % iteration, Sworld, im_rgb)
#import pdb;pdb.set_trace()
# if we need to, make corrections for non-positive depths
#S = util.generate_surface(camera, z_est)
mask = (z_est < INIT_Z)
specular_mask = (L < 0.8)
dark_mask = (L > 0.1)
_mask = np.logical_and(specular_mask, mask)
_mask = np.logical_and(_mask, dark_mask)
# fit reflectance model
r = np.linalg.norm(S, axis=-1)
ndotl = util.calculate_ndotl(camera, S)
falloff, model_params, residual = fit_reflectance_model(
model_type, L[_mask], r.ravel(), r[_mask], ndotl.ravel(),
ndotl[_mask], fit_falloff, camera.width, camera.height,
*extra_params)
#r = np.linalg.norm(S[specular_mask], axis=-1)
#import pdb;pdb.set_trace()
#model_params=np.array([26.15969874,-27.674055,-12.52426,7.579855,21.9768004,24.3911142,-21.7282996,-19.850894,-11.62229,-4.837014])
#model_params=np.array([-19.4837,-490.4796,812.4527,-426.09107,139.2602,351.8061,-388.1591,875.5013,-302.4748,-414.4384])
#falloff = 1.2
#ndotl = util.calculate_ndotl(camera, S)[specular_mask]
#falloff, model_params, residual = fit_reflectance_model(model_type,
# L[specular_mask], r, ndotl, fit_falloff, *extra_params)
#r = np.linalg.norm(S[neighborhood_mask], axis=-1)
#ndotl = util.calculate_ndotl(camera, S)[neighborhood_mask]
#falloff, model_params, residual = fit_reflectance_model(model_type,
# L[neighborhood_mask], r, ndotl, fit_falloff, *extra_params)
# lambda values reflect our confidence in the current surface: 0
# corresponds to only using SFS at a pixel, 1 corresponds to equally
# weighting SFS and the current estimate, and larger values increasingly
# favor using only the current estimate
rdiff = np.abs(r_fixed - get_estimated_r(S, points2D_image))
w = np.log10(r_fixed) - np.log10(rdiff) - np.log10(2.)
lambdas = (np.sum(weights * w[nn_idxs], axis=-1) /
np.sum(weights, axis=-1))
lambdas = np.maximum(lambdas, 0.) # just in case
#
lambdas[~mask] = 0
#if iteration == 0: # don't use current estimated surface on first pass
#lambdas = np.zeros_like(z)
#else:
# r_ratios_postwarp = r_fixed / get_estimated_r(S, points2D_image)
# ratio_diff = np.abs(r_ratios_prewarp - r_ratios_postwarp)
# ratio_diff[ratio_diff == 0] = 1e-10 # arbitrarily high lambda
# feature_lambdas = 1. / ratio_diff
# lambdas = (np.sum(weights * feature_lambdas[nn_idxs], axis=-1) /
# np.sum(weights, axis=-1))
# run SFS
H_lut, dH_lut = compute_H_LUT(model_type, model_params,
NUM_H_LUT_BINS)
#import pdb;pdb.set_trace()
# run SFS
#H_lut = np.ascontiguousarray(H_lut.astype(np.float32))
#dH_lut = np.ascontiguousarray(dH_lut.astype(np.float32))
z = run_sfs(H_lut, dH_lut, camera, L, lambdas, z_est, model_type,
model_params, falloff, vmask,
use_image_weighted_derivatives)
# check for convergence
#diff = np.sum(np.abs(z_old[specular_mask] - z[specular_mask]))
#if diff < CONVERGENCE_THRESHOLD * camera.height * camera.width:
# break
# save the surface
#S = util.generate_surface(camera, z)
#S = (S - image.tvec[np.newaxis,np.newaxis,:]).dot(R)
#util.save_sfs_ply(out_file + '_%i.ply' % iteration, S, im_rgb)
else:
print 'DID NOT CONVERGE'
#import pdb;pdb.set_trace()
S = util.generate_surface(camera, z)
#S = (S - image.tvec[np.newaxis,np.newaxis,:]).dot(R)
util.save_sfs_ply(out_file + '.ply', S, im_rgb)
z.astype(np.float32).tofile(out_file + '_z.bin')
# save the surface
#S = util.generate_surface(camera, z)
#S = (S - image.tvec[np.newaxis,np.newaxis,:]).dot(R)
#S, r_ratios = nearest_neighbor_warp(weights, nn_idxs,
# points2D_image, r_fixed, util.generate_surface(camera, z))
#util.save_sfs_ply(out_file + '_warped.ply', S, im_rgb)
#z = np.maximum(S[:,:,2], 1e-6)
#z.astype(np.float32).tofile(out_file + '_warped_z.bin')
#reflectance_models.save_reflectance_model(out_file + '_reflectance.txt',
# model_name, residual, model_params, falloff, *extra_params)
print
def run(colmap_folder, sfs_results_path, gt_path, out_folder):
#import pdb;pdb.set_trace();
if not sfs_results_path.endswith('/'):
sfs_results_path += '/'
if not gt_path.endswith('/'):
gt_path += '/'
if not out_folder.endswith('/'):
out_folder += '/'
scene_manager = SceneManager(colmap_folder)
scene_manager.load_cameras()
camera = scene_manager.cameras[1] # assume single camera
# initial values on unit sphere
x, y = camera.get_image_grid()
r_scale = np.sqrt(x * x + y * y + 1.)
image = np.empty((camera.height, camera.width, 3))
#image[:,:,0] = 1.
if not os.path.exists(out_folder):
os.mkdir(out_folder)
# iterate through rendered BRDFs
#for brdf_name in os.listdir(sfs_results_path):
for brdf_name in ['phantom']:
brdf_folder = sfs_results_path + brdf_name + '/'
if not os.path.isdir(brdf_folder): continue
with open(out_folder + brdf_name + '.csv', 'w') as f:
print>>f, HEADER1 % brdf_name
print>>f, HEADER2
# iterate through reflectance models
#for rm_name in os.listdir(brdf_folder):
#for model_name in MODELS:
model_folder = brdf_folder + '/'#model_name + '/'
# percent of pixels within 1, 2, 5mm of the GT surface
data = [list() for _ in THRESHOLDS]
###
datasum = []
###
# iterate through SFS surfaces
for filename in os.listdir(model_folder):
if not filename.endswith('_z.bin'): continue
name = filename[:-6]
#import pdb;pdb.set_trace();
util.save_sfs_ply('%s_%s.ply' % (name, 'est'),
np.dstack((x, y, np.ones_like(x))) *
np.fromfile(model_folder + filename, dtype=np.float32).reshape(
camera.height, camera.width, 1))
#
util.save_sfs_ply('%s_%s.ply' % (name, 'gt'),
np.dstack((x, y, np.ones_like(x)))/WORLD_SCALE *
np.fromfile(gt_path + name+'.bin', dtype=np.float32).reshape(
camera.height, camera.width, 1))
#break
r_est = WORLD_SCALE * r_scale * np.fromfile(
model_folder + filename, dtype=np.float32).reshape(
camera.height, camera.width)
r_gt = r_scale * np.fromfile(
gt_path + name + '.bin', dtype=np.float32).reshape(
camera.height, camera.width)
rdiff = np.abs(r_est - r_gt)
#rdiff = np.concatenate((
# rdiff[:,:20].ravel(), rdiff[:,-20:].ravel(),
# rdiff[20:-20,:20].ravel(), rdiff[20:-20,-20:].ravel()))
#rdiff = rdiff[(r_gt > 5.0) & (r_gt < 20.0)]
inv_size = 1. / float(rdiff.size)
for i, t in enumerate(THRESHOLDS):
data[i].append(np.count_nonzero(rdiff < t) * inv_size)
#data[i].append(np.mean((rdiff / r_gt)[rdiff < t]))
#data[i].append(np.mean((r_est / r_gt)))
###
datasum.append(1. - (r_est / r_gt).ravel())
datasum = np.concatenate(datasum)
print np.mean(datasum), np.std(datasum)
###
data = np.array(data)
data = np.mean(data, axis=1), np.std(data, axis=1)
model_name='power'
print>>f, model_name + ',' + ','.join('%f,%f' % d for d in izip(*data))
def run_iterative_sfs(out_file, model_type, fit_falloff, *extra_params):
#if os.path.exists(out_file + '_z.bin'): # don't re-run existing results
# return
if model_type == sfs.LAMBERTIAN_MODEL:
model_name = 'LAMBERTIAN_MODEL'
elif model_type == sfs.OREN_NAYAR_MODEL:
model_name = 'OREN_NAYAR_MODEL'
elif model_type == sfs.PHONG_MODEL:
model_name = 'PHONG_MODEL'
elif model_type == sfs.COOK_TORRANCE_MODEL:
model_name = 'COOK_TORRANCE_MODEL'
elif model_type == sfs.POWER_MODEL:
model_name = 'POWER_MODEL'
if model_type == sfs.POWER_MODEL:
print '%s_%i' % (model_name, extra_params[0])
else:
print model_name
S = S0.copy()
z = S0[:,:,2]
for iteration in xrange(MAX_NUM_ITER):
print 'Iteration', iteration
z_old = z
if get_initial_warp:
#z_gt=util.load_point_ply('C:\\Users\\user\\Documents\\UNC\\Research\\ColonProject\\code\\Rui\\SFS_CPU\\frame0859.jpg_gt.ply')
# warp to 3D points
if iteration>-1:
S, r_ratios = nearest_neighbor_warp(weights, nn_idxs,
points2D_image, r_fixed, util.generate_surface(camera, z))
z_est = np.maximum(S[:,:,2], 1e-6)
S=util.generate_surface(camera, z_est)
else:
z_est=z
S=util.generate_surface(camera, z_est)
#util.save_sfs_ply('warp' + '.ply', S, im_rgb)
#util.save_xyz('test.xyz',points3D);
#z=z_est
#break
#z_est=z
else:
#import pdb;pdb.set_trace()
#z_est = extract_depth_map(camera,ref_surf_name,R,image)
z_est=np.fromfile(
'C:\Users\user\Documents\UNC\Research\ColonProject\code\SFS_Program_from_True\endo_evaluation\gt_surfaces\\frame0859.jpg.bin', dtype=np.float32).reshape(
camera.height, camera.width)/WORLD_SCALE
z_est=z_est.astype(float)
#S, r_ratios = nearest_neighbor_warp(weights, nn_idxs,
# points2D_image, r_fixed, util.generate_surface(camera, z_est))
z_est = np.maximum(z_est[:,:], 1e-6)
#Sworld = (S - image.tvec[np.newaxis,np.newaxis,:]).dot(R)
S = util.generate_surface(camera, z_est)
#util.save_sfs_ply('test' + '.ply', S, im_rgb)
#util.save_sfs_ply(out_file + '_warp_%i.ply' % iteration, Sworld, im_rgb)
#import pdb;pdb.set_trace()
# if we need to, make corrections for non-positive depths
#S = util.generate_surface(camera, z_est)
mask = (z_est < INIT_Z)
specular_mask=(L<0.8)
dark_mask=(L>0.1)
_mask=np.logical_and(specular_mask,mask)
_mask=np.logical_and(_mask,dark_mask)
# fit reflectance model
r = np.linalg.norm(S, axis=-1)
ndotl = util.calculate_ndotl(camera, S)
falloff, model_params, residual = fit_reflectance_model(model_type,
L[_mask],r.ravel(), r[_mask],ndotl.ravel(), ndotl[_mask], fit_falloff,camera.width,camera.height, *extra_params)
#r = np.linalg.norm(S[specular_mask], axis=-1)
#import pdb;pdb.set_trace()
#model_params=np.array([26.15969874,-27.674055,-12.52426,7.579855,21.9768004,24.3911142,-21.7282996,-19.850894,-11.62229,-4.837014])
#model_params=np.array([-19.4837,-490.4796,812.4527,-426.09107,139.2602,351.8061,-388.1591,875.5013,-302.4748,-414.4384])
#falloff = 1.2
#ndotl = util.calculate_ndotl(camera, S)[specular_mask]
#falloff, model_params, residual = fit_reflectance_model(model_type,
# L[specular_mask], r, ndotl, fit_falloff, *extra_params)
#r = np.linalg.norm(S[neighborhood_mask], axis=-1)
#ndotl = util.calculate_ndotl(camera, S)[neighborhood_mask]
#falloff, model_params, residual = fit_reflectance_model(model_type,
# L[neighborhood_mask], r, ndotl, fit_falloff, *extra_params)
# lambda values reflect our confidence in the current surface: 0
# corresponds to only using SFS at a pixel, 1 corresponds to equally
# weighting SFS and the current estimate, and larger values increasingly
# favor using only the current estimate
rdiff = np.abs(r_fixed - get_estimated_r(S, points2D_image))
w = np.log10(r_fixed) - np.log10(rdiff) - np.log10(2.)
lambdas = (np.sum(weights * w[nn_idxs], axis=-1) /
np.sum(weights, axis=-1))
lambdas = np.maximum(lambdas, 0.) # just in case
#
lambdas[~mask] = 0
#if iteration == 0: # don't use current estimated surface on first pass
#lambdas = np.zeros_like(z)
#else:
# r_ratios_postwarp = r_fixed / get_estimated_r(S, points2D_image)
# ratio_diff = np.abs(r_ratios_prewarp - r_ratios_postwarp)
# ratio_diff[ratio_diff == 0] = 1e-10 # arbitrarily high lambda
# feature_lambdas = 1. / ratio_diff
# lambdas = (np.sum(weights * feature_lambdas[nn_idxs], axis=-1) /
# np.sum(weights, axis=-1))
# run SFS
H_lut, dH_lut = compute_H_LUT(model_type, model_params, NUM_H_LUT_BINS)
#import pdb;pdb.set_trace()
# run SFS
#H_lut = np.ascontiguousarray(H_lut.astype(np.float32))
#dH_lut = np.ascontiguousarray(dH_lut.astype(np.float32))
z = run_sfs(H_lut,dH_lut,camera, L, lambdas, z_est, model_type, model_params, falloff,vmask,
use_image_weighted_derivatives)
# check for convergence
#diff = np.sum(np.abs(z_old[specular_mask] - z[specular_mask]))
#if diff < CONVERGENCE_THRESHOLD * camera.height * camera.width:
# break
# save the surface
#S = util.generate_surface(camera, z)
#S = (S - image.tvec[np.newaxis,np.newaxis,:]).dot(R)
#util.save_sfs_ply(out_file + '_%i.ply' % iteration, S, im_rgb)
else:
print 'DID NOT CONVERGE'
#import pdb;pdb.set_trace()
S = util.generate_surface(camera, z)
#S = (S - image.tvec[np.newaxis,np.newaxis,:]).dot(R)
util.save_sfs_ply(out_file + '.ply', S, im_rgb)
z.astype(np.float32).tofile(out_file + '_z.bin')
# save the surface
#S = util.generate_surface(camera, z)
#S = (S - image.tvec[np.newaxis,np.newaxis,:]).dot(R)
#S, r_ratios = nearest_neighbor_warp(weights, nn_idxs,
# points2D_image, r_fixed, util.generate_surface(camera, z))
#util.save_sfs_ply(out_file + '_warped.ply', S, im_rgb)
#z = np.maximum(S[:,:,2], 1e-6)
#z.astype(np.float32).tofile(out_file + '_warped_z.bin')
#reflectance_models.save_reflectance_model(out_file + '_reflectance.txt',
# model_name, residual, model_params, falloff, *extra_params)
print
