def list_of_vectors_of_concatenated_active_joint_gradients(vid, active_joints):
grad = skvp.gradient(vid)
lst = []
for frame in grad.frames:
vec = []
for j in active_joints:
vec.extend(frame[j])
lst.append(np.array(vec))
return lst
def detect_rest_sequences(vid, active_joints, exponent=-1.5):
der = skvp.median(skvp.gradient(vid))
rest_index_counter = {}
for joint in active_joints:
stats = [np.linalg.norm(f[joint])**exponent for f in der.frames]
stats_mean = np.mean(stats)
rest_indices = [i for i, val in enumerate(stats) if val > stats_mean]
margin = int(round(len(der) * 0.075))
rest_indices = [
val for val in rest_indices
if val > margin and val < len(der) - margin
]
for ri in rest_indices:
if ri not in rest_index_counter:
rest_index_counter[ri] = 0
rest_index_counter[ri] += 1
thresh = int(len(active_joints) * 0.3)
rest_indices = sorted(
[ri for ri in rest_index_counter if rest_index_counter[ri] > thresh])
sequences = []
first = None
last = None
for ind in rest_indices:
if first == None:
first = ind
last = ind
continue
if ind - last > len(der) * 0.1:
sequences.append((first, last))
first = ind
last = ind
if last != None:
sequences.append((first, last))
index_trans_func = lambda x: x + 1
return sequences, index_trans_func
def train(input_dir, output_file, warp_mode, output_ref_vid):
vid_files = [
os.path.join(input_dir, fname) for fname in os.listdir(input_dir)
]
print('Loading training videos...')
vids = [skvp.load(path) for path in vid_files]
print('Normalizing training videos...')
vids = [skvp.project_to_body_plane(vid, 0, 4, 8) for vid in vids]
connections_mean_lengths = get_connection_mean_lengths(vids)
vids = [
skvp.scaled_connections(vid, connections_mean_lengths) for vid in vids
]
vids_original_lengths_before_filters = [len(vid) for vid in vids]
vids_mean_len_before_filters = int(
round(np.mean(vids_original_lengths_before_filters)))
print('Scaling training videos to have the same length')
vids = [
skvp.create_length_scaled_video(
vid, num_frames=vids_mean_len_before_filters) for vid in vids
]
print('Applying filters')
vids = [skvp.median(vid) for vid in vids]
vid_pyramids = [skvp.pyramid(vid, [3, 3, 3], [1, 1, 1]) for vid in vids]
vids = [pyr.get_level(1) for pyr in vid_pyramids]
vids_new_len = len(vids[0])
# Saving original videos before tempora alignment
vids_nowrap = [vid[:] for vid in vids]
print('Applying temporal alignment')
ref_vid_index, vids, active_joints, num_rests_in_motion, ref_vid_warping_indices, all_vids_warping_sequences = our_warping_function(
vids)
if warp_mode == 'none':
vids = vids_nowrap
if warp_mode == 'dtw':
vids = warp_using_dtw(vids_nowrap, ref_vid_index, active_joints)
if output_ref_vid != None:
skvp.dump(vids_nowrap[ref_vid_index], output_ref_vid)
# Computing discrete temporal gradients
ders = [skvp.gradient(vid) for vid in vids]
connections = skvp.distinct_connections(vids[0])
# Writing motion metadata into output model file
f = open(output_file, 'w')
f.write('ActiveJoints={0}\n'.format(str(active_joints)))
f.write('VidLength={0:d}\n'.format(vids_mean_len_before_filters))
f.write('NumRests={0:d}\n'.format(num_rests_in_motion))
f.write('WarpIndices={0}\n'.format(ref_vid_warping_indices))
f.write('ConnectionLengths={0}\n'.format(str(connections_mean_lengths)))
f.write('--stats--\n')
print('Calculating stats and writing to output file...')
# Computing and writing time-related statistics - original video lengths and aligned sequence original lengths
mean, dist_mean, dist_std = get_centroid_dist_mean_and_dist_std(
vids_original_lengths_before_filters)
data_unit = {
'mean': mean,
'dist_mean': dist_mean,
'dist_std': dist_std,
'type': 'video_length_frames_raw'
}
f.write('{0}\n'.format(str(data_unit)))
for i in range(len(all_vids_warping_sequences[0]) + 1):
if i == 0:
seq_lengths = [v[0] + 1 for v in all_vids_warping_sequences]
elif i < len(all_vids_warping_sequences[0]):
seq_lengths = [
v[i] - v[i - 1] + 1 for v in all_vids_warping_sequences
]
else:
seq_lengths = [
vids_new_len - v[i - 1] + 1 for v in all_vids_warping_sequences
]
mean, dist_mean, dist_std = get_centroid_dist_mean_and_dist_std(
seq_lengths)
data_unit = {
'mean': mean,
'dist_mean': dist_mean,
'dist_std': dist_std,
'type': 'sequence_length',
'sequence_num': i
}
f.write('{0}\n'.format(str(data_unit)))
# Computing and writing joint-related statistics
for frame_num in range(vids_new_len):
for joint in range(vids[0].get_num_joints()):
points = [vid.frames[frame_num][joint] for vid in vids]
xs = [p[0] for p in points]
ys = [p[1] for p in points]
zs = [p[2] for p in points]
mean, dist_mean, dist_std = get_centroid_dist_mean_and_dist_std(xs)
data_unit = {
'frame': frame_num,
'joint': joint,
'mean': mean,
'dist_mean': dist_mean,
'dist_std': dist_std,
'type': 'location_x'
}
f.write('{0}\n'.format(str(data_unit)))
mean, dist_mean, dist_std = get_centroid_dist_mean_and_dist_std(ys)
data_unit = {
'frame': frame_num,
'joint': joint,
'mean': mean,
'dist_mean': dist_mean,
'dist_std': dist_std,
'type': 'location_y'
}
f.write('{0}\n'.format(str(data_unit)))
mean, dist_mean, dist_std = get_centroid_dist_mean_and_dist_std(zs)
data_unit = {
'frame': frame_num,
'joint': joint,
'mean': mean,
'dist_mean': dist_mean,
'dist_std': dist_std,
'type': 'location_z'
}
f.write('{0}\n'.format(str(data_unit)))
if frame_num < vids_new_len - 1: # We have one frame less in the gradient
der_points = [vid.frames[frame_num][joint] for vid in ders]
xs = [p[0] for p in der_points]
ys = [p[1] for p in der_points]
zs = [p[2] for p in der_points]
mean, dist_mean, dist_std = get_centroid_dist_mean_and_dist_std(
xs)
data_unit = {
'frame': frame_num,
'joint': joint,
'mean': mean,
'dist_mean': dist_mean,
'dist_std': dist_std,
'type': 'gradient_x'
}
f.write('{0}\n'.format(str(data_unit)))
mean, dist_mean, dist_std = get_centroid_dist_mean_and_dist_std(
ys)
data_unit = {
'frame': frame_num,
'joint': joint,
'mean': mean,
'dist_mean': dist_mean,
'dist_std': dist_std,
'type': 'gradient_y'
}
f.write('{0}\n'.format(str(data_unit)))
mean, dist_mean, dist_std = get_centroid_dist_mean_and_dist_std(
zs)
data_unit = {
'frame': frame_num,
'joint': joint,
'mean': mean,
'dist_mean': dist_mean,
'dist_std': dist_std,
'type': 'gradient_z'
}
f.write('{0}\n'.format(str(data_unit)))
for joint_i in range(vids[0].get_num_joints()):
for joint_j in range(joint_i + 1, vids[0].get_num_joints()):
if (joint_i, joint_j) in connections:
# Skipping neighbors!!
continue
dists = [
np.linalg.norm(vid.frames[frame_num][joint_i] -
vid.frames[frame_num][joint_j])
for vid in vids
]
centroid, dist_mean, dist_std = get_centroid_dist_mean_and_dist_std(
dists)
data_unit = {
'frame': frame_num,
'joint_i': joint_i,
'joint_j': joint_j,
'mean': float(centroid),
'dist_mean': dist_mean,
'dist_std': dist_std,
'type': 'joint_distance'
}
f.write('{0}\n'.format(str(data_unit)))
for lower_joint_index, join_index, higher_joint_index in find_all_joint_trios(
vids[0].get_connections()):
angles = []
for vid in vids:
vec_1 = vid.frames[frame_num][lower_joint_index] - vid.frames[
frame_num][join_index]
vec_2 = vid.frames[frame_num][higher_joint_index] - vid.frames[
frame_num][join_index]
theta = np.arccos(
vec_1.dot(vec_2) /
(np.linalg.norm(vec_1) * np.linalg.norm(vec_2)))
angles.append(theta)
centroid, dist_mean, dist_std = get_centroid_dist_mean_and_dist_std(
angles)
data_unit = {
'frame': frame_num,
'joint_trio':
[lower_joint_index, join_index, higher_joint_index],
'mean': float(centroid),
'dist_mean': dist_mean,
'dist_std': dist_std,
'type': 'joint_angles'
}
f.write('{0}\n'.format(str(data_unit)))
f.close()
def test(model_file, input_video_file, warp_mode, ablation, ref_vid_path):
vid = skvp.load(input_video_file)
vid = skvp.project_to_body_plane(vid, 0, 4, 8)
orig_vid_len_frames = len(vid)
vid_len, active_joints, num_rests, warp_indices, connection_lengths, stats = parse_model_file(
model_file)
vid = skvp.scaled_connections(vid, connection_lengths)
vid = skvp.create_length_scaled_video(vid, num_frames=vid_len)
vid = skvp.median(vid)
pyr = skvp.pyramid(vid, [3, 3, 3], [1, 1, 1])
vid = pyr.get_level(1)
rest_sequences = []
exponent = -1.5
ignore_subsequence_lengths = False
while len(rest_sequences) < num_rests:
exponent *= 0.9
if exponent > -0.5:
rest_sequences = [(warp_indices[i - 1], warp_indices[i])
for i in range(len(warp_indices)) if i % 2 == 1]
index_transform_function = lambda x: x + 1
ignore_subsequence_lengths = True
break
rest_sequences, index_transform_function = detect_rest_sequences(
vid, active_joints, exponent)
vec = []
for sq in rest_sequences:
if sq[0] == sq[1]:
sq = (sq[0] - 1, sq[1] + 1)
vec.extend(sq)
vec = [index_transform_function(val) for val in vec]
if warp_mode in ('our'):
vid = warp_video(vid, vec, warp_indices)
elif warp_mode == 'dtw':
refvid = skvp.load(ref_vid_path)
ref_vals = list_of_vectors_of_concatenated_active_joint_gradients(
refvid, active_joints)
vid_vals = list_of_vectors_of_concatenated_active_joint_gradients(
vid, active_joints)
matches, cost, mapping_1, mapping_2, matrix = simpledtw.dtw(
ref_vals, vid_vals)
warped = vid[:1] # Starting from one frame, as grad has n-1 frames
for mapped_indices in mapping_1:
warped.add_frame(vid.frames[mapped_indices[-1]])
vid = warped
der = skvp.gradient(vid)
costs = []
group_weights = {'ActiveJoint': 0.73, 'NonActiveJoint': 0.02, 'Time': 0.25}
if 'active' in ablation:
group_weights = {
'ActiveJoint': 0.15,
'NonActiveJoint': 0.6,
'Time': 0.25
}
if 'time' in ablation:
group_weights['ActiveJoint'] /= (1.0 - group_weights['Time'])
group_weights['NonActiveJoint'] /= (1.0 - group_weights['Time'])
group_weights['Time'] = 0
for stat in stats:
if stat['type'] == 'sequence_length':
stat['start_frame'] = (vec[stat['sequence_num'] - 1] +
1) if stat['sequence_num'] > 0 else 0
stat['end_frame'] = vec[
stat['sequence_num']] if stat['sequence_num'] < len(
vec) else len(vid) - 1
if ignore_subsequence_lengths and stat['type'] == 'sequence_length':
print('Ignoring seq lengths')
continue
if 'joint' in stat or 'joint_trio' in stat:
stat['type_group'] = 'NonActiveJoint'
elif 'joint_i' in stat:
stat[
'type_group'] = 'NoGroupIgnoreMe' if 'active' not in ablation else 'NonActiveJoint'
else:
stat['type_group'] = 'Time'
if 'joint' in stat and stat['joint'] in active_joints:
stat['type_group'] = 'ActiveJoint'
if 'joint_i' in stat and stat[
'joint_i'] in active_joints and 'joint_j' in stat and stat[
'joint_j'] in active_joints:
stat['type_group'] = 'ActiveJoint'
if 'joint_trio' in stat:
for j in stat['joint_trio']:
if j in active_joints:
stat['type_group'] = 'ActiveJoint'
if stat['type'] == 'location_x':
val = vid.frames[stat['frame']][stat['joint']][0]
elif stat['type'] == 'location_y':
val = vid.frames[stat['frame']][stat['joint']][1]
elif stat['type'] == 'location_z':
val = vid.frames[stat['frame']][stat['joint']][2]
elif stat['type'] == 'gradient_x':
val = der.frames[stat['frame']][stat['joint']][0]
elif stat['type'] == 'gradient_y':
val = der.frames[stat['frame']][stat['joint']][1]
elif stat['type'] == 'gradient_z':
val = der.frames[stat['frame']][stat['joint']][2]
elif stat['type'] == 'joint_distance':
val = np.linalg.norm(vid.frames[stat['frame']][stat['joint_i']] -
vid.frames[stat['frame']][stat['joint_j']])
elif stat['type'] == 'joint_angles':
joint_trio = stat['joint_trio']
vec_1 = vid.frames[stat['frame']][joint_trio[0]] - vid.frames[
stat['frame']][joint_trio[1]]
vec_2 = vid.frames[stat['frame']][joint_trio[2]] - vid.frames[
stat['frame']][joint_trio[1]]
val = np.arccos(
vec_1.dot(vec_2) /
(np.linalg.norm(vec_1) * np.linalg.norm(vec_2)))
elif stat['type'] == 'video_length_frames_raw':
val = orig_vid_len_frames
elif stat['type'] == 'sequence_length':
if stat['sequence_num'] == 0:
val = vec[0] + 1
elif stat['sequence_num'] < len(vec):
val = vec[stat['sequence_num']] - vec[stat['sequence_num'] -
1] + 1
else:
val = len(vid) - vec[stat['sequence_num'] - 1] + 1
else:
continue
# Avoiding division by 0 when we normalize the distance
val = round(val, 7)
stat['mean'] = round(stat['mean'], 7)
stat['dist_std'] = round(stat['dist_std'], 7)
stat['dist_mean'] = round(stat['dist_mean'], 7)
try:
dist = abs(val - stat['mean'])
except:
print('stat is: {0}'.format(str(stat)))
print('val is: {0}'.format(str(val)))
print('mean is: {0}'.format(str(stat['mean'])))
if stat['dist_std'] == 0:
if dist - stat['dist_mean'] == 0:
dist_in_stds = 0
else:
dist_in_stds = np.inf
else:
dist_in_stds = (dist - stat['dist_mean']) / stat['dist_std']
stat['cost'] = dist_in_stds
stat['test_val'] = val
costed_stats = [s for s in stats
if 'cost' in s] # and s['type'] == 'joint_angles']
costed_active_joint_stats = [
s for s in costed_stats if s['type_group'] == 'ActiveJoint'
]
costed_nonactive_joint_stats = [
s for s in costed_stats if s['type_group'] == 'NonActiveJoint'
]
costed_time_stats = [s for s in costed_stats if s['type_group'] == 'Time']
lambdaa = 2
group_to_stats = {
'Time': costed_time_stats,
'ActiveJoint': costed_active_joint_stats,
'NonActiveJoint': costed_nonactive_joint_stats
}
feedback_items, avg_active_bad_segment_length, avg_param_num_active_segments, avg_param_num_nonactive_segments = produce_feedback(
costed_stats, group_weights, ablation)
scorelambda = 2.9
scorelambdatime = 7.5
for item in feedback_items:
# Here we duplicate the weight manipulation, just to be able to sort the feedbacks
# The same manupulation will be applied on real items (depending on ablation settings) - we don't apply ablation on feedback
item['cost'] *= group_weights[item['type_group']]
item['cost'] /= (scorelambdatime
if item['type_group'] == 'Time' else scorelambda)
item['cost'] /= len(
[s['cost'] for s in group_to_stats[item['type_group']]])
if item['type_group'] == 'ActiveJoint':
item['cost'] *= (0.75**avg_param_num_active_segments)
feedback_items.sort(key=lambda x: x['cost'], reverse=True)
if len(feedback_items) > 5:
feedback_items = feedback_items[:5]
last_cost = None
for i, fi in enumerate(feedback_items):
if last_cost == None or last_cost < 2 * fi['cost']:
last_cost = fi['cost']
continue
# Removing irrelevant feedback
feedback_items = feedback_items[:i]
break
feedback_items = [generate_feedback_text(fi, 15) for fi in feedback_items]
if len(feedback_items) == 0:
print('FEEDBACK: empty! congrats :)')
for feedback_item in feedback_items:
print('FEEDBACK: ' + str(feedback_item))
if 'diminish' not in ablation and 'active' not in ablation:
for s in costed_active_joint_stats:
s['cost'] *= (0.75**avg_param_num_active_segments)
#Vector scores
active_joint_score = max(
0, 1.0 - abs(
sum(s['cost']
for s in costed_active_joint_stats if s['cost'] > 0) /
(scorelambda * len([s['cost']
for s in costed_active_joint_stats]))))
nonactive_joint_score = max(
0, 1.0 - abs(
sum(s['cost']
for s in costed_nonactive_joint_stats if s['cost'] > 0) /
(scorelambda *
len([s['cost'] for s in costed_nonactive_joint_stats]))))
time_score = max(
0, 1.0 - abs(
sum(s['cost'] for s in costed_time_stats if s['cost'] > 2.5) /
(scorelambdatime * len([s['cost'] for s in costed_time_stats]))))
print('Scores: Active: {0:f}, NonActive: {1:f}, Time: {2:f}'.format(
active_joint_score, nonactive_joint_score, time_score))
final_score = active_joint_score * group_weights[
'ActiveJoint'] + nonactive_joint_score * group_weights[
'NonActiveJoint'] + time_score * group_weights['Time']
print('Score: {0:f}'.format(final_score))