def make_tdpw():
root = '/globalwork/datasets/3DPW'
body_joint_names = (
'pelv,lhip,rhip,bell,lkne,rkne,spin,lank,rank,thor,ltoe,rtoe,neck,lcla,rcla,head,lsho,'
'rsho,lelb,relb,lwri,rwri,lhan,rhan'.split(','))
selected_joints = [*range(1, 24), 0]
joint_names = [body_joint_names[j] for j in selected_joints]
edges = ('head-neck-lcla-lsho-lelb-lwri-lhan,'
'neck-rcla-rsho-relb-rwri-rhan,'
'neck-thor-spin-bell-pelv-lhip-lkne-lank-ltoe,'
'pelv-rhip-rkne-rank-rtoe')
joint_info = p3ds.JointInfo(joint_names, edges)
def get_examples(phase, pool):
result = []
seq_filepaths = glob.glob(f'{root}/sequenceFiles/{phase}/*.pkl')
for filepath in seq_filepaths:
with open(filepath, 'rb') as f:
seq = pickle.load(f, encoding='latin1')
seq_name = seq['sequence']
intrinsics = seq['cam_intrinsics']
extrinsics_per_frame = seq['cam_poses']
for i_person, (coord_seq, coords2d_seq, trans_seq, camvalid_seq) in enumerate(zip(
seq['jointPositions'], seq['poses2d'], seq['trans'], seq['campose_valid'])):
for i_frame, (coords, coords2d, trans, extrinsics, campose_valid) in enumerate(
zip(coord_seq, coords2d_seq, trans_seq, extrinsics_per_frame,
camvalid_seq)):
if not campose_valid or np.all(coords2d == 0):
continue
impath = f'{root}/imageFiles/{seq_name}/image_{i_frame:05d}.jpg'
camera = cameralib.Camera(
extrinsic_matrix=extrinsics, intrinsic_matrix=intrinsics,
world_up=(0, 1, 0))
camera.t *= 1000
world_coords = (coords.reshape(-1, 3))[selected_joints] * 1000
camera2 = cameralib.Camera(intrinsic_matrix=intrinsics, world_up=(0, -1, 0))
camcoords = camera.world_to_camera(world_coords)
imcoords = camera.world_to_image(world_coords)
bbox = boxlib.expand(boxlib.bb_of_points(imcoords), 1.15)
ex = p3ds.Pose3DExample(impath, camcoords, bbox=bbox, camera=camera2)
noext, ext = os.path.splitext(os.path.relpath(impath, root))
new_image_relpath = f'tdpw_downscaled/{noext}_{i_person:03d}.jpg'
pool.apply_async(
make_efficient_example,
(ex, new_image_relpath, 1, False, "2021-07-09T12:28:07"),
callback=result.append)
return result
with util.BoundedPool(None, 120) as pool:
train_examples = get_examples('train', pool)
val_examples = get_examples('validation', pool)
test_examples = get_examples('test', pool)
test_examples = [*train_examples, *val_examples, *test_examples]
test_examples.sort(key=lambda ex: ex.image_path)
return p3ds.Pose3DDataset(joint_info, None, None, test_examples)
def make_joint_info():
all_joint_names = (
'thor,spi4,spi2,spin,pelv,neck,head,htop,lcla,lsho,lelb,lwri,lhan,rcla,rsho,relb,rwri,'
'rhan,lhip,lkne,lank,lfoo,ltoe,rhip,rkne,rank,rfoo,rtoe'.split(','))
selected_joints = [
0, 1, 2, 8, 12, 13, 17, 21, 22, 26, 27, 7, 5, 14, 15, 16, 9, 10, 11, 23, 24, 25, 18, 19, 20,
3, 6, 4]
joint_names = [all_joint_names[j] for j in selected_joints]
edges = (
'htop-head-neck-lcla-lsho-lelb-lwri-lhan,neck-rcla-rsho-relb-rwri-rhan,'
'neck-spi4-thor-spi2-spin-pelv-lhip-lkne-lank-lfoo-ltoe,pelv-rhip-rkne-rank-rfoo-rtoe')
joint_info = p3ds.JointInfo(joint_names, edges)
return joint_info, selected_joints
def make_mupots():
joint_names = (
'htop,neck,rsho,relb,rwri,lsho,lelb,lwri,rhip,rkne,rank,lhip,lkne,lank,spin,head,pelv'
)
edges = ('htop-head-neck-spin-pelv-lhip-lkne-lank,'
'lwri-lelb-lsho-neck-rsho-relb-rwri,pelv-rhip-rkne-rank')
joint_info = p3ds.JointInfo(joint_names, edges)
#import data.muco
#joint_info = data.muco.make_joint_info()[0]
root = f'{paths.DATA_ROOT}/mupots'
intrinsic_matrices = util.load_json(f'{root}/camera_intrinsics.json')
dummy_coords = np.ones((joint_info.n_joints, 3))
detections_all = util.load_pickle(f'{root}/yolov3_detections.pkl')
examples_val = []
examples_test = []
for i_seq in range(1, 21):
annotations = matlabfile.load(
f'{root}/TS{i_seq}/annot.mat')['annotations']
intrinsic_matrix = intrinsic_matrices[f'TS{i_seq}']
camera = cameralib.Camera(np.zeros(3),
np.eye(3),
intrinsic_matrix,
distortion_coeffs=None,
world_up=(0, -1, 0))
n_frames = annotations.shape[0]
for i_frame in range(n_frames):
image_relpath = f'TS{i_seq}/img_{i_frame:06d}.jpg'
detections_frame = detections_all[image_relpath]
image_path = f'{root}/{image_relpath}'
for detection in detections_frame:
confidence = detection[4]
if confidence > 0.1:
ex = p3ds.Pose3DExample(os.path.relpath(
image_path, paths.DATA_ROOT),
dummy_coords,
detection[:4],
camera,
mask=None,
univ_coords=dummy_coords,
scene_name=f'TS{i_seq}')
examples_test.append(ex)
return p3ds.Pose3DDataset(joint_info,
valid_examples=examples_val,
test_examples=examples_test)
def generate_poseviz_gt(i_subject, activity_name, camera_id):
camera_names = ['54138969', '55011271', '58860488', '60457274']
camera_name = camera_names[camera_id]
image_relpaths, world_coords_all, bboxes, camera = get_examples(
i_subject, activity_name, camera_id, frame_step=1, correct_S9=True)
results = []
examples = []
for image_relpath, world_coords, bbox in zip(image_relpaths,
world_coords_all, bboxes):
results.append({
'gt_poses': [world_coords.tolist()],
'camera_intrinsics':
camera.intrinsic_matrix.tolist(),
'camera_extrinsics':
camera.get_extrinsic_matrix().tolist(),
'image_path':
image_relpath,
'bboxes': [bbox.tolist()]
})
ex = ps3d.Pose3DExample(image_relpath,
world_coords,
bbox,
camera,
activity_name=activity_name)
examples.append(ex)
joint_names = ('rhip,rkne,rank,lhip,lkne,lank,tors,neck,head,htop,'
'lsho,lelb,lwri,rsho,relb,rwri,pelv'.split(','))
edges = ('htop-head-neck-lsho-lelb-lwri,neck-rsho-relb-rwri,'
'neck-tors-pelv-lhip-lkne-lank,pelv-rhip-rkne-rank')
joint_info = ps3d.JointInfo(joint_names, edges)
ds = ps3d.Pose3DDataset(joint_info, test_examples=examples)
util.dump_pickle(
ds,
f'{paths.DATA_ROOT}/h36m/poseviz/S{i_subject}_{activity_name}_{camera_name}.pkl'
)
output = {}
output['joint_names'] = joint_info.names
output['stick_figure_edges'] = joint_info.stick_figure_edges
output['world_up'] = camera.world_up.tolist()
output['frame_infos'] = results
util.dump_json(
output,
f'{paths.DATA_ROOT}/h36m/poseviz/S{i_subject}_{activity_name}_{camera_name}.json'
)
def make_mpi_inf_3dhp(camera_ids=(0, 1, 2, 4, 5, 6, 7, 8)):
all_short_names = (
'spi3,spi4,spi2,spin,pelv,neck,head,htop,lcla,lsho,lelb,lwri,lhan,rcla,rsho,relb,rwri,'
'rhan,lhip,lkne,lank,lfoo,ltoe,rhip,rkne,rank,rfoo,rtoe'.split(','))
test_set_selected_joints = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 15, 16, 14]
selected_joints = [7, 5, 14, 15, 16, 9, 10, 11, 23, 24, 25, 18, 19, 20, 3, 6, 4]
joint_names = [all_short_names[j] for j in selected_joints]
edges = ('htop-head-neck-lsho-lelb-lwri,neck-rsho-relb-rwri,neck-spin-pelv-lhip-lkne-lank,'
'pelv-rhip-rkne-rank')
joint_info = p3ds.JointInfo(joint_names, edges)
root_3dhp = f'{paths.DATA_ROOT}/3dhp'
detections_all = util.load_pickle(f'{paths.DATA_ROOT}/3dhp/yolov3_person_detections.pkl')
#################################
# TRAINING AND VALIDATION SET
#################################
num_frames = np.asarray(
[[6416, 12430], [6502, 6081], [12488, 12283], [6171, 6675], [12820, 12312], [6188, 6145],
[6239, 6320], [6468, 6054]])
train_subjects = [0, 1, 2, 3, 4, 5, 6]
valid_subjects = [7] # this is my own arbitrary split for validation (Istvan Sarandi)
train_examples = []
valid_examples = []
pool = util.BoundedPool(None, 120)
for i_subject, i_seq, i_cam in itertools.product(
train_subjects + valid_subjects, range(2), camera_ids):
seqpath = f'{root_3dhp}/S{i_subject + 1}/Seq{i_seq + 1}'
print(f'Processing {seqpath} camera {i_cam}')
cam3d_coords = [ann.reshape([ann.shape[0], -1, 3])[:, selected_joints]
for ann in matlabfile.load(f'{seqpath}/annot.mat')['annot3']]
univ_cam3d_coords = [ann.reshape([ann.shape[0], -1, 3])[:, selected_joints]
for ann in matlabfile.load(f'{seqpath}/annot.mat')['univ_annot3']]
cameras = load_cameras(f'{seqpath}/camera.calibration')
examples_container = train_examples if i_subject in train_subjects else valid_examples
frame_step = 5
prev_coords = None
camera = cameras[i_cam]
n_frames = num_frames[i_subject, i_seq]
if i_subject == 5 and i_seq == 1 and i_cam == 2:
# This video is shorter for some reason
n_frames = 3911
for i_frame in util.progressbar(range(0, n_frames, frame_step)):
image_relpath = (
f'3dhp/S{i_subject + 1}/Seq{i_seq + 1}/'
f'imageSequence/img_{i_cam}_{i_frame:06d}.jpg')
cam_coords = cam3d_coords[i_cam][i_frame]
world_coords = cameras[i_cam].camera_to_world(cam_coords)
univ_camcoords = univ_cam3d_coords[i_cam][i_frame]
univ_world_coords = cameras[i_cam].camera_to_world(univ_camcoords)
# Check if the joints are within the image frame bounds
if not np.all(camera.is_visible(world_coords, [2048, 2048])):
continue
im_coords = camera.camera_to_image(cam_coords)
bbox = get_bbox(im_coords, image_relpath, detections_all)
# Adaptive temporal sampling
if (prev_coords is not None and
np.all(np.linalg.norm(world_coords - prev_coords, axis=1) < 100)):
continue
prev_coords = world_coords
mask_path = image_relpath.replace('imageSequence', 'FGmasks')
new_image_relpath = image_relpath.replace('3dhp', '3dhp_downscaled')
ex = p3ds.Pose3DExample(
image_relpath, world_coords, bbox, camera, mask=mask_path,
univ_coords=univ_world_coords)
pool.apply_async(make_efficient_example, (ex, new_image_relpath, 1, True),
callback=examples_container.append)
print('Waiting for tasks...')
pool.close()
pool.join()
print('Done...')
#################################
# TEST SET
#################################
test_examples = []
cam1_4 = make_3dhp_test_camera(
sensor_size=np.array([10, 10]), im_size=np.array([2048, 2048]), focal_length=7.32506,
pixel_aspect=1.00044, center_offset=np.array([-0.0322884, 0.0929296]), distortion=None,
origin=np.array([3427.28, 1387.86, 309.42]), up=np.array([-0.208215, 0.976233, 0.06014]),
right=np.array([0.000575281, 0.0616098, -0.9981]))
cam5_6 = make_3dhp_test_camera(
sensor_size=np.array([10, 5.625]), im_size=np.array([1920, 1080]), focal_length=8.770747185,
pixel_aspect=0.993236423, center_offset=np.array([-0.104908645, 0.104899704]),
distortion=np.array([-0.276859611, 0.131125256, -0.000360494, -0.001149441, -0.049318332]),
origin=np.array([-2104.3074, 1038.6707, -4596.6367]),
up=np.array([0.025272345, 0.995038509, 0.096227370]),
right=np.array([-0.939647257, -0.009210289, 0.342020929]))
activity_names = [
'Stand/Walk', 'Exercise', 'Sit on Chair', 'Reach/Crouch', 'On Floor', 'Sports', 'Misc.']
for i_subject in range(1, 7):
seqpath = f'{root_3dhp}/TS{i_subject}'
annotation_path = f'{seqpath}/annot_data.mat'
with h5py.File(annotation_path, 'r') as m:
cam3d_coords = np.array(m['annot3'])[:, 0, test_set_selected_joints]
univ_cam3d_coords = np.array(m['univ_annot3'])[:, 0, test_set_selected_joints]
valid_frames = np.where(m['valid_frame'][:, 0])[0]
activity_ids = m['activity_annotation'][:, 0].astype(int) - 1
camera = cam1_4 if i_subject <= 4 else cam5_6
scene = ['green-screen', 'no-green-screen', 'outdoor'][(i_subject - 1) // 2]
for i_frame in valid_frames:
image_relpath = f'3dhp/TS{i_subject}/imageSequence/img_{i_frame + 1:06d}.jpg'
cam_coords = cam3d_coords[i_frame]
univ_camcoords = univ_cam3d_coords[i_frame]
activity = activity_names[activity_ids[i_frame]]
world_coords = camera.camera_to_world(cam_coords)
univ_world_coords = camera.camera_to_world(univ_camcoords)
im_coords = camera.camera_to_image(cam_coords)
bbox = get_bbox(im_coords, image_relpath, detections_all)
ex = p3ds.Pose3DExample(
image_relpath, world_coords, bbox, camera, activity_name=activity,
scene_name=scene, univ_coords=univ_world_coords)
test_examples.append(ex)
train_examples.sort(key=lambda x: x.image_path)
valid_examples.sort(key=lambda x: x.image_path)
test_examples.sort(key=lambda x: x.image_path)
return p3ds.Pose3DDataset(joint_info, train_examples, valid_examples, test_examples)
def make_h36m(train_subjects=(1, 5, 6, 7, 8),
valid_subjects=(),
test_subjects=(9, 11),
correct_S9=True,
partial_visibility=False):
joint_names = ('rhip,rkne,rank,lhip,lkne,lank,tors,neck,head,htop,'
'lsho,lelb,lwri,rsho,relb,rwri,pelv'.split(','))
edges = ('htop-head-neck-lsho-lelb-lwri,neck-rsho-relb-rwri,'
'neck-tors-pelv-lhip-lkne-lank,pelv-rhip-rkne-rank')
joint_info = ps3d.JointInfo(joint_names, edges)
if not util.all_disjoint(train_subjects, valid_subjects, test_subjects):
raise Exception('Set of train, val and test subject must be disjoint.')
# use last subject of the non-test subjects for validation
train_examples = []
test_examples = []
valid_examples = []
pool = util.BoundedPool(None, 120)
if partial_visibility:
dir_suffix = '_partial'
further_expansion_factor = 1.8
else:
dir_suffix = '' if correct_S9 else 'incorrect_S9'
further_expansion_factor = 1
for i_subject in [*test_subjects, *train_subjects, *valid_subjects]:
if i_subject in train_subjects:
examples_container = train_examples
elif i_subject in valid_subjects:
examples_container = valid_examples
else:
examples_container = test_examples
frame_step = 5 if i_subject in train_subjects else 64
for activity_name, camera_id in itertools.product(
get_activity_names(i_subject), range(4)):
print(f'Processing S{i_subject} {activity_name} {camera_id}')
image_relpaths, world_coords_all, bboxes, camera = get_examples(
i_subject,
activity_name,
camera_id,
frame_step=frame_step,
correct_S9=correct_S9)
prev_coords = None
for image_relpath, world_coords, bbox in zip(
util.progressbar(image_relpaths), world_coords_all,
bboxes):
# Using very similar examples is wasteful when training. Therefore:
# skip frame if all keypoints are within a distance compared to last stored frame.
# This is not done when testing, as it would change the results.
if (i_subject in train_subjects and prev_coords is not None
and np.all(
np.linalg.norm(world_coords -
prev_coords, axis=1) < 100)):
continue
prev_coords = world_coords
activity_name = activity_name.split(' ')[0]
ex = ps3d.Pose3DExample(image_relpath,
world_coords,
bbox,
camera,
activity_name=activity_name)
new_image_relpath = image_relpath.replace(
'h36m', f'h36m_downscaled{dir_suffix}')
pool.apply_async(
make_efficient_example,
(ex, new_image_relpath, further_expansion_factor),
callback=examples_container.append)
print('Waiting for tasks...')
pool.close()
pool.join()
print('Done...')
train_examples.sort(key=lambda x: x.image_path)
valid_examples.sort(key=lambda x: x.image_path)
test_examples.sort(key=lambda x: x.image_path)
return ps3d.Pose3DDataset(joint_info, train_examples, valid_examples,
test_examples)
def make_mupots_yolo():
all_short_names = (
'thor,spi4,spi2,spin,pelv,neck,head,htop,lcla,lsho,lelb,lwri,lhan,rcla,rsho,relb,rwri,'
'rhan,lhip,lkne,lank,lfoo,ltoe,rhip,rkne,rank,rfoo,rtoe'.split(','))
# originally: [7, 5, 14, 15, 16, 9, 10, 11, 23, 24, 25, 18, 19, 20, 4, 3, 6]
selected_joints = [
7, 5, 14, 15, 16, 9, 10, 11, 23, 24, 25, 18, 19, 20, 3, 6, 4
]
order_joints = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 15, 16, 14]
joint_names = [all_short_names[j] for j in selected_joints]
j = p3ds.JointInfo.make_id_map(joint_names)
edges = [(j.htop, j.head), (j.head, j.neck), (j.neck, j.lsho),
(j.lsho, j.lelb), (j.lelb, j.lwri), (j.neck, j.rsho),
(j.rsho, j.relb), (j.relb, j.rwri), (j.neck, j.spin),
(j.spin, j.pelv), (j.pelv, j.lhip), (j.lhip, j.lkne),
(j.lkne, j.lank), (j.pelv, j.rhip), (j.rhip, j.rkne),
(j.rkne, j.rank)]
joint_info = p3ds.JointInfo(j, edges)
root = f'{paths.DATA_ROOT}/mupots'
intrinsic_matrices = util.load_json(f'{root}/camera_intrinsics.json')
dummy_coords = np.ones((joint_info.n_joints, 3))
detections_all = util.load_pickle(f'{root}/yolov3_detections.pkl')
examples_val = []
examples_test = []
for i_seq in range(1, 21):
annotations = matlabfile.load(
f'{root}/TS{i_seq}/annot.mat')['annotations']
intrinsic_matrix = intrinsic_matrices[f'TS{i_seq}']
camera = cameralib.Camera(np.zeros(3),
np.eye(3),
intrinsic_matrix,
distortion_coeffs=None,
world_up=(0, -1, 0))
n_people = annotations.shape[1]
n_frames = annotations.shape[0]
for i_frame in range(n_frames):
image_relpath = f'TS{i_seq}/img_{i_frame:06d}.jpg'
detections_frame = detections_all[image_relpath]
image_path = f'{root}/{image_relpath}'
for detection in detections_frame:
if detection[4] > 0.1:
ex = p3ds.Pose3DExample(image_path,
dummy_coords,
detection[:4],
camera,
mask=None,
univ_coords=dummy_coords,
scene_name=f'TS{i_seq}')
examples_test.append(ex)
gt_people = []
for i_person in range(n_people):
world_coords = np.array(
annotations[i_frame, i_person].annot3.T[order_joints],
dtype=np.float32)
univ_world_coords = np.array(
annotations[i_frame, i_person].univ_annot3.T[order_joints],
dtype=np.float32)
im_coords = camera.world_to_image(world_coords)
gt_box = boxlib.expand(boxlib.bb_of_points(im_coords), 1.1)
ex = p3ds.Pose3DExample(image_path,
world_coords,
gt_box,
camera,
mask=None,
univ_coords=univ_world_coords,
scene_name=f'TS{i_seq}')
gt_people.append(ex)
confident_detections = [
det for det in detections_frame if det[-1] > 0.1
]
if confident_detections:
iou_matrix = np.array([[
boxlib.iou(gt_person.bbox, box[:4])
for box in confident_detections
] for gt_person in gt_people])
gt_indices, detection_indices = scipy.optimize.linear_sum_assignment(
-iou_matrix)
for i_gt, i_det in zip(gt_indices, detection_indices):
if iou_matrix[i_gt, i_det] > 0.1:
ex = gt_people[i_gt]
ex.bbox = np.array(confident_detections[i_det][:4])
examples_val.append(ex)
return p3ds.Pose3DDataset(joint_info,
valid_examples=examples_val,
test_examples=examples_test)