def plotUsers(image, users):
# embed()
# if type(users) == dict:
# users = [u for u in users.values()]
# for u in users:
# uvw = [-1]
for u in users.keys():
if users[u]['tracked']:
xyz = users[u]['com']
uvw = skel2depth(np.array([xyz]), image.shape)[0]
''' Only plot if there are valid coordinates (not [0,0,0])'''
if uvw[0] > 0:
if users[u]['tracked'] and len(
users[u]['jointPositions'].keys()) > 0:
'''Colorize COM'''
cv2.rectangle(image, tuple([uvw[0] - 3, uvw[1] - 3]),
tuple([uvw[0] + 3, uvw[1] + 3]), (4000))
'''Plot skeleton'''
pts = [j for j in users[u]['jointPositions'].values()]
skel = skel2depth(np.array(pts), image.shape)
from pyKinectTools.utils.SkeletonUtils import display_skeletons
image = display_skeletons(image,
skel,
color=(image.max(), 0, 0),
skel_type='Kinect')
def learn_msr_video(name, offsets_1, offsets_2): ''' ''' depth_file = name + "depth.bin" color_file = name + "rgb.avi" skeleton_file = name + "skeleton.txt" ''' Read data from each video/sequence ''' try: depthIms, maskIms = read_MSR_depth_ims(depth_file) depthIms *= maskIms depthIms /= 10 colorIms = read_MSR_color_ims(color_file) skels_world, _ = read_MSR_skeletons(skeleton_file) skels_world[:,2]/=10 except: print "Error getting frame features" return -1,-1 all_features = [] all_labels = [] for i in xrange(0, len(depthIms), 25): im_depth = depthIms[i] skel_pos = skel2depth(skels_world[i], rez=[240,320]) # ''' --- NOTE THIS 10 PX OFFSET IN THE MSR DATASET !!! --- ''' # skel_pos[:,0] -= 10 features, pixel_labels = learn_frame(im_depth, skel_pos, offsets_1, offsets_2) ''' Stack features ''' all_features += [features] all_labels += [pixel_labels] return np.concatenate(all_features), np.concatenate(all_labels)
def learn_msr_video(name, offsets_1, offsets_2):
'''
'''
depth_file = name + "depth.bin"
color_file = name + "rgb.avi"
skeleton_file = name + "skeleton.txt"
''' Read data from each video/sequence '''
try:
depthIms, maskIms = read_MSR_depth_ims(depth_file)
depthIms *= maskIms
depthIms /= 10
colorIms = read_MSR_color_ims(color_file)
skels_world, _ = read_MSR_skeletons(skeleton_file)
skels_world[:, 2] /= 10
except:
print "Error getting frame features"
return -1, -1
all_features = []
all_labels = []
for i in xrange(0, len(depthIms), 25):
im_depth = depthIms[i]
skel_pos = skel2depth(skels_world[i], rez=[240, 320])
# ''' --- NOTE THIS 10 PX OFFSET IN THE MSR DATASET !!! --- '''
# skel_pos[:,0] -= 10
features, pixel_labels = learn_frame(im_depth, skel_pos, offsets_1,
offsets_2)
''' Stack features '''
all_features += [features]
all_labels += [pixel_labels]
return np.concatenate(all_features), np.concatenate(all_labels)
def transform_skels(skels, transformation, output='image'):
'''
---Parameters---
skels : list of skeletons in frame 1
transformation : 4x4 transform from frame 1 to 2
output : 'image' or 'world' for either coordinate system
---Result---
skels_out : skeletons in frame 2
'''
skels_out = []
for skel_c1 in skels:
if np.all(skel_c1 != -1):
skels_mask = skel_c1 == 0
# Convert to depth coordinate system
skel_c1 = depth2world(skel2depth(skel_c1, [240, 320]), [240, 320])
# Transform from cam1 -> cam2
skel_c2 = np.dot(transformation[:3, :3],
skel_c1.T).T + transformation[:3, 3]
if len(skel_c2) != N_MSR_JOINTS:
skel_c2 = kinect_to_msr_skel(skel_c2)
skel_c2[skels_mask] = 0
if output == 'world':
skels_out += [skel_c2]
elif output == 'image':
# Get skel in image (cam2) coordinates
skel_im2 = world2depth(skel_c2, [240, 320])
skels_out += [skel_im2]
return skels_out
def transform_skels(skels, transformation, output='image'):
'''
---Parameters---
skels : list of skeletons in frame 1
transformation : 4x4 transform from frame 1 to 2
output : 'image' or 'world' for either coordinate system
---Result---
skels_out : skeletons in frame 2
'''
skels_out = []
for skel_c1 in skels:
if np.all(skel_c1 != -1):
skels_mask = skel_c1 == 0
# Convert to depth coordinate system
skel_c1 = depth2world(skel2depth(skel_c1, [240,320]), [240,320])
# Transform from cam1 -> cam2
skel_c2 = np.dot(transformation[:3,:3], skel_c1.T).T + transformation[:3,3]
if len(skel_c2) != N_MSR_JOINTS:
skel_c2 = kinect_to_msr_skel(skel_c2)
skel_c2[skels_mask] = 0
if output=='world':
skels_out += [skel_c2]
elif output=='image':
# Get skel in image (cam2) coordinates
skel_im2 = world2depth(skel_c2, [240,320])
skels_out += [skel_im2]
return skels_out
def plotUsers(image, users):
# embed()
# if type(users) == dict:
# users = [u for u in users.values()]
# for u in users:
# uvw = [-1]
for u in users.keys():
if users[u]['tracked']:
xyz = users[u]['com']
uvw = skel2depth(np.array([xyz]), image.shape)[0]
''' Only plot if there are valid coordinates (not [0,0,0])'''
if uvw[0] > 0:
if users[u]['tracked'] and len(users[u]['jointPositions'].keys()) > 0:
'''Colorize COM'''
cv2.rectangle(image, tuple([uvw[0]-3, uvw[1]-3]), tuple([uvw[0]+3, uvw[1]+3]), (4000))
'''Plot skeleton'''
pts = [j for j in users[u]['jointPositions'].values()]
skel = skel2depth(np.array(pts), image.shape)
from pyKinectTools.utils.SkeletonUtils import display_skeletons
image = display_skeletons(image, skel, color=(image.max(),0,0), skel_type='Kinect')
def main(visualize=True):
n_cameras = 1
cam = KinectPlayer(base_dir='./',
device=1,
bg_subtraction=True,
get_depth=True,
get_color=True,
get_skeleton=False,
fill_images=False)
if n_cameras == 2:
cam2 = KinectPlayer(base_dir='./',
device=2,
bg_subtraction=True,
get_depth=True,
get_color=get_color,
get_skeleton=get_skel,
fill_images=fill)
# Transformation matrix from first to second camera
try:
data = pickle.load(open("Registration.dat", 'r'))
transform_c1_to_c2 = data['transform']
transform = True
except:
transform = False
pass
current_frame = 0
all_joint_ims_z = []
all_joint_ims_c = []
framerate = 1
while cam.next():
# print "Frame ", current_frame
# Update frames
if n_cameras == 2:
cam2.next()
# cam2.sync_cameras(cam)
current_frame += 1
if current_frame % framerate != 0:
# current_frame += 1
continue
# Transform skels from cam1 to cam2
# cam_skels = [np.array(cam.users[s]['jointPositions'].values()) for s in cam.users.keys()]
# cam_skels = [np.array(cam.users[s]['jointPositions'].values()) for s in cam.users]
# Get rid of bad skels
# cam_skels = [s for s in cam_skels if np.all(s[0] != -1)]
# if len(cam_skels) == 0:
# continue
# Save images
if 0:
joint_ims_z = []
joint_ims_c = []
dx = 10
skel_tmp = skel2depth(cam_skels[0], [240, 320])
for j_pos in skel_tmp:
# embed()
joint_ims_z += [
cam.depthIm[j_pos[0] - dx:j_pos[0] + dx,
j_pos[1] - dx:j_pos[1] + dx]
]
joint_ims_c += [
cam.colorIm[j_pos[0] - dx:j_pos[0] + dx,
j_pos[1] - dx:j_pos[1] + dx]
]
if len(joint_ims_z) > 0:
all_joint_ims_z += [joint_ims_z]
all_joint_ims_c += [joint_ims_c]
if 1:
# if transform:
# cam2_skels = transform_skels(cam_skels, transform_c1_to_c2, 'image')
# try:
# depth = cam2.get_person()
# if learn:
# rf.add_frame(depth, cam2_skels[0])
# else:
# rf.infer_pose(depth)
# except:
# pass
if visualize:
# cam2.depthIm = display_skeletons(cam2.depthIm, cam2_skels[0], (5000,), skel_type='Low')
# skel1 = kinect_to_msr_skel(skel2depth(cam_skels[0], [240,320]))
# cam.depthIm = display_skeletons(cam.depthIm, skel1, (5000,), skel_type='Low')
# embed()
cam.visualize(color=True,
depth=True,
text=True,
colorize=True,
depth_bounds=[500, 3500])
if n_cameras == 2:
cam2.visualize(color=True, depth=True)
embed()
print 'Done'
def main(visualize=False, learn=False):
# Init both cameras
# fill = True
fill = False
get_color = True
cam = KinectPlayer(base_dir='./', device=1, bg_subtraction=True, get_depth=True, get_color=get_color, get_skeleton=True, fill_images=fill)
cam2 = KinectPlayer(base_dir='./', device=2, bg_subtraction=True, get_depth=True, get_color=get_color, get_skeleton=True, fill_images=fill)
# Transformation matrix from first to second camera
data = pickle.load(open("Registration.dat", 'r'))
transform_c1_to_c2 = data['transform']
# Get random forest parameters
if learn:
rf = RFPose(offset_max=100, feature_count=300)
else:
rf = RFPose()
rf.load_forest()
ii = 0
# cam.next(200)
all_joint_ims_z = []
all_joint_ims_c = []
while cam.next() and ii < 200:
# Update frames
cam2.sync_cameras(cam)
if ii%10 != 0:
ii += 1
continue
# Transform skels from cam1 to cam2
cam_skels = [np.array(cam.users[s]['jointPositions'].values()) for s in cam.users.keys()]
# Get rid of bad skels
cam_skels = [s for s in cam_skels if np.all(s[0] != -1)]
if len(cam_skels) == 0:
continue
ii+=1
# Save images
if 1:
joint_ims_z = []
joint_ims_c = []
dx = 10
skel_tmp = skel2depth(cam_skels[0], [240,320])
for j_pos in skel_tmp:
# embed()
joint_ims_z += [cam.depthIm[j_pos[0]-dx:j_pos[0]+dx, j_pos[1]-dx:j_pos[1]+dx]]
joint_ims_c += [cam.colorIm[j_pos[0]-dx:j_pos[0]+dx, j_pos[1]-dx:j_pos[1]+dx]]
if len(joint_ims_z) > 0:
all_joint_ims_z += [joint_ims_z]
all_joint_ims_c += [joint_ims_c]
if 0:
cam2_skels = transform_skels(cam_skels, transform_c1_to_c2, 'image')
try:
depth = cam2.get_person()
if learn:
rf.add_frame(depth, cam2_skels[0])
else:
rf.infer_pose(depth)
if visualize:
cam2.depthIm = display_skeletons(cam2.depthIm, cam2_skels[0], (5000,), skel_type='Low')
skel1 = kinect_to_msr_skel(skel2depth(cam_skels[0], [240,320]))
cam.depthIm = display_skeletons(cam.depthIm, skel1, (5000,), skel_type='Low')
cam.visualize()
cam2.visualize()
except:
pass
embed()
if learn:
print "Starting forest"
rf.learn_forest()
print 'Done'
def main_infer(rf_name=None):
'''
'''
if rf_name is None:
import os
files = os.listdir('Saved_Params/')
rf_name = 'Saved_Params/' + files[-1]
print "Classifier file:",rf_name
# Load classifier data
data = pickle.load(open(rf_name))
rf = data['rf']
offsets_1, offsets_2 = data['offsets']
''' Read data from each video/sequence '''
depthIms = []
skels_world = []
if 1:
# VP = KinectPlayer(base_dir='/Users/colin/Data/Office_25Feb2013/', device=2, get_depth=True, get_color=False, get_skeleton=True, get_mask=False)
VP = KinectPlayer(base_dir='/Users/colin/Data/Room_close_26Feb13/', device=1, get_depth=True, get_color=False, get_skeleton=True, get_mask=False)
_,_ = VP.get_n_skeletons(50)
depthIms, skels_world = VP.get_n_skeletons(100)
# depthIms = np.array(depthIms)[:,:,::-1]
else:
# name = 'a01_s01_e02_'
name = 'a01_s10_e02_'
# name = 'a02_s06_e02_'
# name = 'a05_s02_e02_'
depth_file = name + "depth.bin"
color_file = name + "rgb.avi"
skeleton_file = name + "skeleton.txt"
try:
depthIms, maskIms = read_MSR_depth_ims(depth_file)
depthIms *= maskIms
depthIms /= 10
# colorIms = read_MSR_color_ims(color_file)
skels_world, skels_im = read_MSR_skeletons(skeleton_file)
skels_world[:,2]/=10
except:
print "Error reading data"
''' Process data '''
all_pred_ims = []
for i in xrange(1, len(depthIms), 1):
# try:
if 1:
print i
''' Get frame data '''
im_depth = depthIms[i]
# im_depth[160:, :] = 0
skel_pos = skel2depth(skels_world[i], rez=[240,320])
# ''' --- NOTE THIS 10 PX OFFSET IN THE MSR DATASET !!! --- '''
# skel_pos[:,0] -= 10
skel_pos_pred, im_predict = infer_pose(im_depth, rf, offsets_1, offsets_2)
# Overlay skeletons
if 1:
# colorIm = colorIms[i]
# im_predict = colorIm
cv2.imshow("forest_prediction", im_predict/float(im_predict.max()))
im_predict = np.repeat(im_depth[:,:,None].astype(np.float), 3, -1)
# embed()
im_predict[im_depth>0] -= im_depth[im_depth>0].min()
im_predict /= float(im_predict.max()/255.)
im_predict = im_predict.astype(np.uint8)
# im_predict = np.repeat(im_predict[:,:,None], 3, -1)
# im_predict /= float(im_predict.max())*255
# im_predict = im_predict.astype(np.uint8)
im_predict = display_skeletons(im_predict, skel_pos, (255,0,0), SKEL_DISPLAY_MODE)
im_predict = display_skeletons(im_predict, skel_pos_pred, (0,255,0), SKEL_DISPLAY_MODE)
im_predict = display_skeletons(im_predict, skel_pos, (255,0,0), SKEL_DISPLAY_MODE)
im_predict = display_skeletons(im_predict, skel_pos_pred, (0,255,0), SKEL_DISPLAY_MODE)
# embed()
# max_ = (im_predict * (im_predict < 255)).max()
all_pred_ims += [im_predict]
''' Visualize '''
if 1:
cv2.putText(im_predict, "Blue=Truth", (10, 210), cv2.FONT_HERSHEY_DUPLEX, .5, (int(im_predict.max()/2), 0, 0))
cv2.putText(im_predict, "Green=Predict", (10, 230), cv2.FONT_HERSHEY_DUPLEX, .5, (0, int(im_predict.max()/2), 0))
cv2.imshow("prediction", im_predict)
ret = cv2.waitKey(10)
if ret > 0: break
time.sleep(.5)
# except:
# print "Frame failed:", i
# break
embed()
def main_learn():
'''
'''
offsets_1, offsets_2 = create_rf_offsets()
all_features = []
all_labels = []
''' My data '''
if 1:
# VP = KinectPlayer(base_dir='/Users/colin/Data/Office_25Feb2013/', device=2, get_depth=True, get_color=False, get_skeleton=True, get_mask=False)
# VP = KinectPlayer(base_dir='/Users/colin/Data/Room_close_26Feb13/', device=1, get_depth=True, get_color=False, get_skeleton=True, get_mask=False)
# Scrap first 50 frames
# depthIms, skels_world = VP.get_n_skeletons(50)
depthIms, skels_world = VP.get_n_skeletons(400)
# depthIms, skels_world = VP.get_n_skeletons(-1)
print "{0:d} frames loaded".format(len(depthIms))
# for i in range(0, len(depthIms), 1):
for i in range(0, 40):
# try:
if 1:
print "Frame", i
depthIms, skels_world = VP.get_n_skeletons(10)
if depthIms == -1:
break
depth = depthIms[0]
skel_w = skels_world[0]
# depth = depthIms[i]
# skel_w = skels_world[i]
skel_im = skel2depth(skel_w, rez=[240,320])
features, pixel_labels = learn_frame(depth, skel_im, offsets_1, offsets_2)
all_features += [features]
all_labels += [pixel_labels]
# except:
# print "Error in frame {0:d}".format(i)
del depthIms, skels_world
''' MSR data'''
if 0:
names = create_MSR_filenames(np.arange(1)+1, np.arange(3, 5)+1, [2])
# Parallelize feature collection
from joblib import Parallel, delayed
if 1:
for n_set in chunks(names, 1):
print "Computing with one thread. Current feature count:", len(all_features)
all_data = learn_msr_video(n_set[0], offsets_1, offsets_2)
if all_features == []:
all_features = all_data[0]
all_labels = all_data[1]
else:
all_features = np.vstack([all_features, all_data[0]])
all_labels = np.hstack([all_labels, all_data[1]])
else:
for n_set in chunks(names, 100):
print "Computing with multiple threads. Current feature count:", len(all_features)
all_data = Parallel(n_jobs=-1, verbose=True, pre_dispatch=2)( delayed(learn_frame)(n, offsets_1, offsets_2) for n in n_set )
# Account for bad frames
all_features += [f[0] for f in all_data if f[0] is not -1]
all_labels += [f[1] for f in all_data if f[1] is not -1]
print "Done computing this set of features/labels"
print "Done computing all features/labels"
all_features = np.vstack(all_features)
all_labels = np.hstack(all_labels)
print "Starting forest"
rf = RFClassifier(n_estimators=6,
criterion='entropy',\
max_depth=20,
max_features='auto',\
oob_score=False,\
n_jobs=-1,
random_state=None,
verbose=1,\
min_samples_leaf=1)
rf.fit(all_features, all_labels)
save_data("Saved_Params/"+str(time.time())+"_forests.dat", {'rf':rf, 'offsets':[offsets_1, offsets_2]})
def main_infer(rf_name=None):
'''
'''
if rf_name is None:
import os
files = os.listdir('Saved_Params/')
rf_name = 'Saved_Params/' + files[-1]
print "Classifier file:", rf_name
# Load classifier data
data = pickle.load(open(rf_name))
rf = data['rf']
offsets_1, offsets_2 = data['offsets']
''' Read data from each video/sequence '''
depthIms = []
skels_world = []
if 1:
# VP = KinectPlayer(base_dir='/Users/colin/Data/Office_25Feb2013/', device=2, get_depth=True, get_color=False, get_skeleton=True, get_mask=False)
VP = KinectPlayer(base_dir='/Users/colin/Data/Room_close_26Feb13/',
device=1,
get_depth=True,
get_color=False,
get_skeleton=True,
get_mask=False)
_, _ = VP.get_n_skeletons(50)
depthIms, skels_world = VP.get_n_skeletons(100)
# depthIms = np.array(depthIms)[:,:,::-1]
else:
# name = 'a01_s01_e02_'
name = 'a01_s10_e02_'
# name = 'a02_s06_e02_'
# name = 'a05_s02_e02_'
depth_file = name + "depth.bin"
color_file = name + "rgb.avi"
skeleton_file = name + "skeleton.txt"
try:
depthIms, maskIms = read_MSR_depth_ims(depth_file)
depthIms *= maskIms
depthIms /= 10
# colorIms = read_MSR_color_ims(color_file)
skels_world, skels_im = read_MSR_skeletons(skeleton_file)
skels_world[:, 2] /= 10
except:
print "Error reading data"
''' Process data '''
all_pred_ims = []
for i in xrange(1, len(depthIms), 1):
# try:
if 1:
print i
''' Get frame data '''
im_depth = depthIms[i]
# im_depth[160:, :] = 0
skel_pos = skel2depth(skels_world[i], rez=[240, 320])
# ''' --- NOTE THIS 10 PX OFFSET IN THE MSR DATASET !!! --- '''
# skel_pos[:,0] -= 10
skel_pos_pred, im_predict = infer_pose(im_depth, rf, offsets_1,
offsets_2)
# Overlay skeletons
if 1:
# colorIm = colorIms[i]
# im_predict = colorIm
cv2.imshow("forest_prediction",
im_predict / float(im_predict.max()))
im_predict = np.repeat(im_depth[:, :, None].astype(np.float),
3, -1)
# embed()
im_predict[im_depth > 0] -= im_depth[im_depth > 0].min()
im_predict /= float(im_predict.max() / 255.)
im_predict = im_predict.astype(np.uint8)
# im_predict = np.repeat(im_predict[:,:,None], 3, -1)
# im_predict /= float(im_predict.max())*255
# im_predict = im_predict.astype(np.uint8)
im_predict = display_skeletons(im_predict, skel_pos,
(255, 0, 0), SKEL_DISPLAY_MODE)
im_predict = display_skeletons(im_predict, skel_pos_pred,
(0, 255, 0), SKEL_DISPLAY_MODE)
im_predict = display_skeletons(im_predict, skel_pos,
(255, 0, 0), SKEL_DISPLAY_MODE)
im_predict = display_skeletons(im_predict, skel_pos_pred,
(0, 255, 0), SKEL_DISPLAY_MODE)
# embed()
# max_ = (im_predict * (im_predict < 255)).max()
all_pred_ims += [im_predict]
''' Visualize '''
if 1:
cv2.putText(im_predict, "Blue=Truth", (10, 210),
cv2.FONT_HERSHEY_DUPLEX, .5,
(int(im_predict.max() / 2), 0, 0))
cv2.putText(im_predict, "Green=Predict", (10, 230),
cv2.FONT_HERSHEY_DUPLEX, .5,
(0, int(im_predict.max() / 2), 0))
cv2.imshow("prediction", im_predict)
ret = cv2.waitKey(10)
if ret > 0: break
time.sleep(.5)
# except:
# print "Frame failed:", i
# break
embed()
def main_learn():
'''
'''
offsets_1, offsets_2 = create_rf_offsets()
all_features = []
all_labels = []
''' My data '''
if 1:
# VP = KinectPlayer(base_dir='/Users/colin/Data/Office_25Feb2013/', device=2, get_depth=True, get_color=False, get_skeleton=True, get_mask=False)
# VP = KinectPlayer(base_dir='/Users/colin/Data/Room_close_26Feb13/', device=1, get_depth=True, get_color=False, get_skeleton=True, get_mask=False)
# Scrap first 50 frames
# depthIms, skels_world = VP.get_n_skeletons(50)
depthIms, skels_world = VP.get_n_skeletons(400)
# depthIms, skels_world = VP.get_n_skeletons(-1)
print "{0:d} frames loaded".format(len(depthIms))
# for i in range(0, len(depthIms), 1):
for i in range(0, 40):
# try:
if 1:
print "Frame", i
depthIms, skels_world = VP.get_n_skeletons(10)
if depthIms == -1:
break
depth = depthIms[0]
skel_w = skels_world[0]
# depth = depthIms[i]
# skel_w = skels_world[i]
skel_im = skel2depth(skel_w, rez=[240, 320])
features, pixel_labels = learn_frame(depth, skel_im, offsets_1,
offsets_2)
all_features += [features]
all_labels += [pixel_labels]
# except:
# print "Error in frame {0:d}".format(i)
del depthIms, skels_world
''' MSR data'''
if 0:
names = create_MSR_filenames(
np.arange(1) + 1,
np.arange(3, 5) + 1, [2])
# Parallelize feature collection
from joblib import Parallel, delayed
if 1:
for n_set in chunks(names, 1):
print "Computing with one thread. Current feature count:", len(
all_features)
all_data = learn_msr_video(n_set[0], offsets_1, offsets_2)
if all_features == []:
all_features = all_data[0]
all_labels = all_data[1]
else:
all_features = np.vstack([all_features, all_data[0]])
all_labels = np.hstack([all_labels, all_data[1]])
else:
for n_set in chunks(names, 100):
print "Computing with multiple threads. Current feature count:", len(
all_features)
all_data = Parallel(n_jobs=-1, verbose=True, pre_dispatch=2)(
delayed(learn_frame)(n, offsets_1, offsets_2)
for n in n_set)
# Account for bad frames
all_features += [f[0] for f in all_data if f[0] is not -1]
all_labels += [f[1] for f in all_data if f[1] is not -1]
print "Done computing this set of features/labels"
print "Done computing all features/labels"
all_features = np.vstack(all_features)
all_labels = np.hstack(all_labels)
print "Starting forest"
rf = RFClassifier(n_estimators=6,
criterion='entropy',\
max_depth=20,
max_features='auto',\
oob_score=False,\
n_jobs=-1,
random_state=None,
verbose=1,\
min_samples_leaf=1)
rf.fit(all_features, all_labels)
save_data("Saved_Params/" + str(time.time()) + "_forests.dat", {
'rf': rf,
'offsets': [offsets_1, offsets_2]
})
def main(visualize=False, learn=False):
# Init both cameras
# fill = True
fill = False
get_color = True
cam = KinectPlayer(base_dir='./',
device=1,
bg_subtraction=True,
get_depth=True,
get_color=get_color,
get_skeleton=True,
fill_images=fill)
cam2 = KinectPlayer(base_dir='./',
device=2,
bg_subtraction=True,
get_depth=True,
get_color=get_color,
get_skeleton=True,
fill_images=fill)
# Transformation matrix from first to second camera
data = pickle.load(open("Registration.dat", 'r'))
transform_c1_to_c2 = data['transform']
# Get random forest parameters
if learn:
rf = RFPose(offset_max=100, feature_count=300)
else:
rf = RFPose()
rf.load_forest()
ii = 0
# cam.next(200)
all_joint_ims_z = []
all_joint_ims_c = []
while cam.next() and ii < 200:
# Update frames
cam2.sync_cameras(cam)
if ii % 10 != 0:
ii += 1
continue
# Transform skels from cam1 to cam2
cam_skels = [
np.array(cam.users[s]['jointPositions'].values())
for s in cam.users.keys()
]
# Get rid of bad skels
cam_skels = [s for s in cam_skels if np.all(s[0] != -1)]
if len(cam_skels) == 0:
continue
ii += 1
# Save images
if 1:
joint_ims_z = []
joint_ims_c = []
dx = 10
skel_tmp = skel2depth(cam_skels[0], [240, 320])
for j_pos in skel_tmp:
# embed()
joint_ims_z += [
cam.depthIm[j_pos[0] - dx:j_pos[0] + dx,
j_pos[1] - dx:j_pos[1] + dx]
]
joint_ims_c += [
cam.colorIm[j_pos[0] - dx:j_pos[0] + dx,
j_pos[1] - dx:j_pos[1] + dx]
]
if len(joint_ims_z) > 0:
all_joint_ims_z += [joint_ims_z]
all_joint_ims_c += [joint_ims_c]
if 0:
cam2_skels = transform_skels(cam_skels, transform_c1_to_c2,
'image')
try:
depth = cam2.get_person()
if learn:
rf.add_frame(depth, cam2_skels[0])
else:
rf.infer_pose(depth)
if visualize:
cam2.depthIm = display_skeletons(cam2.depthIm,
cam2_skels[0], (5000, ),
skel_type='Low')
skel1 = kinect_to_msr_skel(
skel2depth(cam_skels[0], [240, 320]))
cam.depthIm = display_skeletons(cam.depthIm,
skel1, (5000, ),
skel_type='Low')
cam.visualize()
cam2.visualize()
except:
pass
embed()
if learn:
print "Starting forest"
rf.learn_forest()
print 'Done'
