Various performance improvements.
This commit is contained in:
Daniel
BIN
media/2d-k.png
BIN
media/2d-k.png
Binary file not shown.
|
Before Width: | Height: | Size: 1.7 MiB After Width: | Height: | Size: 1.9 MiB |
BIN
media/2d-p.png
BIN
media/2d-p.png
Binary file not shown.
|
Before Width: | Height: | Size: 1.7 MiB After Width: | Height: | Size: 1.9 MiB |
BIN
media/3d-p.png
BIN
media/3d-p.png
Binary file not shown.
|
Before Width: | Height: | Size: 534 KiB After Width: | Height: | Size: 541 KiB |
2264
media/RESULTS.md
2264
media/RESULTS.md
File diff suppressed because it is too large
Load Diff
@ -352,13 +352,24 @@ def main():
|
||||
time_2d = time.time() - start
|
||||
print("2D time:", time_2d)
|
||||
|
||||
minscores = {
|
||||
# Choose this depending on the fraction of invalid/missing persons
|
||||
# A higher value reduces the number of proposals
|
||||
"panoptic": 0.95,
|
||||
"human36m": 0.96,
|
||||
"mvor": 0.87,
|
||||
"campus": 0.96,
|
||||
"shelf": 0.96,
|
||||
}
|
||||
minscore = minscores.get(dataset_use, 0.95)
|
||||
|
||||
start = time.time()
|
||||
if sum(np.sum(p) for p in poses_2d) == 0:
|
||||
poses3D = np.zeros([1, len(joint_names_3d), 4])
|
||||
poses2D = np.zeros([len(images_2d), 1, len(joint_names_3d), 3])
|
||||
else:
|
||||
poses3D = triangulate_poses.get_3d_pose(
|
||||
poses_2d, label["cameras"], roomparams, joint_names_2d
|
||||
poses_2d, label["cameras"], roomparams, joint_names_2d, minscore
|
||||
)
|
||||
poses2D = []
|
||||
for cam in label["cameras"]:
|
||||
|
||||
@ -288,17 +288,27 @@ def filter_poses(poses3D, poses2D, roomparams, joint_names, drop_few_limbs=True)
|
||||
drop.append(i)
|
||||
continue
|
||||
|
||||
# Drop persons with too small average limb length
|
||||
# Drop persons with too small or high average limb length
|
||||
total_length = 0
|
||||
total_limbs = 0
|
||||
for limb in main_limbs:
|
||||
start_idx = joint_names.index(limb[0])
|
||||
end_idx = joint_names.index(limb[1])
|
||||
if pose[start_idx, -1] < 0.1 or pose[end_idx, -1] < 0.1:
|
||||
continue
|
||||
limb_length = np.linalg.norm(pose[end_idx, :3] - pose[start_idx, :3])
|
||||
total_length += limb_length
|
||||
average_length = total_length / len(main_limbs)
|
||||
total_limbs += 1
|
||||
if total_limbs == 0:
|
||||
drop.append(i)
|
||||
continue
|
||||
average_length = total_length / total_limbs
|
||||
if average_length < 0.1:
|
||||
drop.append(i)
|
||||
continue
|
||||
if average_length > 0.5:
|
||||
drop.append(i)
|
||||
continue
|
||||
|
||||
new_poses3D = []
|
||||
new_poses2D = [[] for _ in range(len(poses2D))]
|
||||
|
||||
@ -28,16 +28,17 @@ core_joints = [
|
||||
|
||||
|
||||
def undistort_points(points: np.ndarray, caminfo: dict):
|
||||
K = np.array(caminfo["K"], dtype=np.float32)
|
||||
DC = np.array(caminfo["DC"][0:5], dtype=np.float32)
|
||||
w = caminfo["width"]
|
||||
h = caminfo["height"]
|
||||
"""Undistorts 2D pixel coordinates"""
|
||||
|
||||
K = np.asarray(caminfo["K"], dtype=np.float32)
|
||||
DC = np.asarray(caminfo["DC"][0:5], dtype=np.float32)
|
||||
|
||||
# Undistort camera matrix
|
||||
w = caminfo["width"]
|
||||
h = caminfo["height"]
|
||||
newK, _ = cv2.getOptimalNewCameraMatrix(K, DC, (w, h), 1, (w, h))
|
||||
caminfo = copy.deepcopy(caminfo)
|
||||
caminfo["K"] = newK.tolist()
|
||||
caminfo["DC"] = [0.0, 0.0, 0.0, 0.0, 0.0]
|
||||
caminfo["K"] = newK
|
||||
caminfo["DC"] = np.array([0.0, 0.0, 0.0, 0.0, 0.0])
|
||||
|
||||
# Undistort points
|
||||
pshape = points.shape
|
||||
@ -45,7 +46,7 @@ def undistort_points(points: np.ndarray, caminfo: dict):
|
||||
points = cv2.undistortPoints(points, K, DC, P=newK)
|
||||
points = points.reshape(pshape)
|
||||
|
||||
return points, caminfo
|
||||
return points
|
||||
|
||||
|
||||
# ==================================================================================================
|
||||
@ -54,10 +55,11 @@ def undistort_points(points: np.ndarray, caminfo: dict):
|
||||
def get_camera_P(cam):
|
||||
"""Calculate opencv-style projection matrix"""
|
||||
|
||||
R = np.array(cam["R"])
|
||||
T = np.array(cam["T"])
|
||||
R = np.asarray(cam["R"])
|
||||
T = np.asarray(cam["T"])
|
||||
K = np.asarray(cam["K"])
|
||||
Tr = R @ (T * -1)
|
||||
P = cam["K"] @ np.hstack([R, Tr])
|
||||
P = K @ np.hstack([R, Tr])
|
||||
return P
|
||||
|
||||
|
||||
@ -69,8 +71,8 @@ def calc_pair_score(pair, poses_2d, camparams, roomparams, joint_names_2d, use_j
|
||||
|
||||
cam1 = camparams[pair[0][0]]
|
||||
cam2 = camparams[pair[0][1]]
|
||||
pose1 = np.array(poses_2d[pair[0][0]][pair[0][2]])
|
||||
pose2 = np.array(poses_2d[pair[0][1]][pair[0][3]])
|
||||
pose1 = poses_2d[pair[0][0]][pair[0][2]]
|
||||
pose2 = poses_2d[pair[0][1]][pair[0][3]]
|
||||
|
||||
# Select core joints
|
||||
jids = [joint_names_2d.index(j) for j in use_joints]
|
||||
@ -88,12 +90,13 @@ def calc_pose_scored(pose1, pose2, cam1, cam2, roomparams):
|
||||
"""Triangulates a pair of persons and scores them based on the reprojection error"""
|
||||
|
||||
# Mask out invisible joints
|
||||
mask1a = pose1[:, 2] >= 0.1
|
||||
mask2a = pose2[:, 2] >= 0.1
|
||||
min_score = 0.1
|
||||
mask1a = pose1[:, 2] >= min_score
|
||||
mask2a = pose2[:, 2] >= min_score
|
||||
mask = mask1a & mask2a
|
||||
|
||||
# If no joints are visible return a low score
|
||||
if np.sum(mask) == 0:
|
||||
# If too few joints are visible return a low score
|
||||
if np.sum(mask) < 3:
|
||||
pose3d = np.zeros([len(pose1), 4])
|
||||
score = 0.0
|
||||
return pose3d, score
|
||||
@ -104,8 +107,7 @@ def calc_pose_scored(pose1, pose2, cam1, cam2, roomparams):
|
||||
P1 = get_camera_P(cam1)
|
||||
P2 = get_camera_P(cam2)
|
||||
points3d = cv2.triangulatePoints(P1, P2, points1, points2)
|
||||
points3d = points3d / points3d[3, :]
|
||||
points3d = points3d[0:3, :].T
|
||||
points3d = (points3d / points3d[3, :])[0:3, :].T
|
||||
pose3d = np.zeros([len(pose1), 4])
|
||||
pose3d[mask] = np.concatenate([points3d, np.ones([points3d.shape[0], 1])], axis=-1)
|
||||
|
||||
@ -115,8 +117,11 @@ def calc_pose_scored(pose1, pose2, cam1, cam2, roomparams):
|
||||
maxs = np.max(pose3d[mask][:, 0:3], axis=0)
|
||||
rsize = np.array(roomparams["room_size"]) / 2
|
||||
rcent = np.array(roomparams["room_center"])
|
||||
wdist = 0.1
|
||||
center_outside = np.any((mean > rsize + rcent) | (mean < -rsize + rcent))
|
||||
limb_outside = np.any((maxs > rsize + rcent + 0.1) | (mins < -rsize + rcent - 0.1))
|
||||
limb_outside = np.any(
|
||||
(maxs > rsize + rcent + wdist) | (mins < -rsize + rcent - wdist)
|
||||
)
|
||||
if center_outside or limb_outside:
|
||||
pose3d[:, 3] = 0.001
|
||||
score = 0.001
|
||||
@ -124,29 +129,33 @@ def calc_pose_scored(pose1, pose2, cam1, cam2, roomparams):
|
||||
|
||||
# Calculate reprojection error
|
||||
poses_3d = np.expand_dims(pose3d, axis=0)
|
||||
repro1, _ = utils_pose.project_poses(poses_3d, cam1)
|
||||
repro2, _ = utils_pose.project_poses(poses_3d, cam2)
|
||||
repro1 = repro1[0]
|
||||
repro2 = repro2[0]
|
||||
error1 = np.linalg.norm(pose1[mask, 0:2] - repro1[mask, 0:2], axis=1)
|
||||
error2 = np.linalg.norm(pose2[mask, 0:2] - repro2[mask, 0:2], axis=1)
|
||||
repro1, dists1 = utils_pose.project_poses(poses_3d, cam1, calc_dists=True)
|
||||
repro2, dists2 = utils_pose.project_poses(poses_3d, cam2, calc_dists=True)
|
||||
error1 = np.linalg.norm(pose1[mask, 0:2] - repro1[0, mask, 0:2], axis=1)
|
||||
error2 = np.linalg.norm(pose2[mask, 0:2] - repro2[0, mask, 0:2], axis=1)
|
||||
|
||||
# Set errors of invisible reprojections to a high value
|
||||
penalty = (cam1["width"] + cam1["height"]) / 2
|
||||
mask1b = (repro1[:, 2] < 0.1)[mask]
|
||||
mask2b = (repro2[:, 2] < 0.1)[mask]
|
||||
mask1b = (repro1[0, :, 2] < min_score)[mask]
|
||||
mask2b = (repro2[0, :, 2] < min_score)[mask]
|
||||
error1[mask1b] = penalty
|
||||
error2[mask2b] = penalty
|
||||
|
||||
# Convert errors to a score
|
||||
scale = (cam1["width"] + cam1["height"]) / 2
|
||||
error1 = error1.clip(0, scale / 4)
|
||||
error2 = error2.clip(0, scale / 4)
|
||||
error1 = error1 / scale
|
||||
error2 = error2 / scale
|
||||
# Scale by image size and distance to the camera
|
||||
iscale = (cam1["width"] + cam1["height"]) / 2
|
||||
error1 = error1.clip(0, iscale / 4) / iscale
|
||||
error2 = error2.clip(0, iscale / 4) / iscale
|
||||
dscale1 = np.sqrt(np.mean(dists1[0, mask]) / 3.5)
|
||||
dscale2 = np.sqrt(np.mean(dists2[0, mask]) / 3.5)
|
||||
error1 = error1 * dscale1
|
||||
error2 = error2 * dscale2
|
||||
error = (error1 + error2) / 2
|
||||
scores = 1.0 / (1.0 + error * 10)
|
||||
score = np.mean(scores)
|
||||
|
||||
# Drop lowest scores
|
||||
drop_k = math.floor(len(pose1) * 0.2)
|
||||
score = np.mean(np.sort(scores, axis=-1)[drop_k:])
|
||||
|
||||
# Add score to 3D pose
|
||||
full_scores = np.zeros([poses_3d.shape[1], 1])
|
||||
@ -159,11 +168,10 @@ def calc_pose_scored(pose1, pose2, cam1, cam2, roomparams):
|
||||
# ==================================================================================================
|
||||
|
||||
|
||||
def calc_grouping(all_pairs):
|
||||
def calc_grouping(all_pairs, min_score: float):
|
||||
"""Groups pairs that share a person"""
|
||||
|
||||
# Calculate the pose center for each pair
|
||||
min_score = 0.9
|
||||
for i in range(len(all_pairs)):
|
||||
pair = all_pairs[i]
|
||||
pose_3d = pair[2][0]
|
||||
@ -177,7 +185,7 @@ def calc_grouping(all_pairs):
|
||||
|
||||
# Create new group if non exists
|
||||
if len(groups) == 0:
|
||||
groups.append([pair[4], pair[2][0], [pair]])
|
||||
groups.append([pair[3], pair[2][0], [pair]])
|
||||
continue
|
||||
|
||||
# Check if the pair matches to an existing group
|
||||
@ -188,7 +196,7 @@ def calc_grouping(all_pairs):
|
||||
for j in range(len(groups)):
|
||||
g0 = groups[j]
|
||||
center0 = g0[0]
|
||||
center1 = pair[4]
|
||||
center1 = pair[3]
|
||||
if np.linalg.norm(center0 - center1) < max_center_dist:
|
||||
pose0 = g0[1]
|
||||
pose1 = pair[2][0]
|
||||
@ -206,14 +214,14 @@ def calc_grouping(all_pairs):
|
||||
if best_group >= 0:
|
||||
# Add pair to existing group and update the mean positions
|
||||
group = groups[best_group]
|
||||
new_center = (group[0] * len(group[1]) + pair[4]) / (len(group[1]) + 1)
|
||||
new_center = (group[0] * len(group[1]) + pair[3]) / (len(group[1]) + 1)
|
||||
new_pose = (group[1] * len(group[1]) + pair[2][0]) / (len(group[1]) + 1)
|
||||
group[2].append(pair)
|
||||
group[0] = new_center
|
||||
group[1] = new_pose
|
||||
else:
|
||||
# Create new group if no match was found
|
||||
groups.append([pair[4], pair[2][0], [pair]])
|
||||
groups.append([pair[3], pair[2][0], [pair]])
|
||||
|
||||
return groups
|
||||
|
||||
@ -221,26 +229,38 @@ def calc_grouping(all_pairs):
|
||||
# ==================================================================================================
|
||||
|
||||
|
||||
def merge_group(poses_3d: np.ndarray):
|
||||
def merge_group(poses_3d: np.ndarray, min_score: float):
|
||||
"""Merges a group of poses into a single pose"""
|
||||
|
||||
# Merge poses to create initial pose
|
||||
# Use only those triangulations with a high score
|
||||
min_score = 0.9
|
||||
mask = poses_3d[:, :, 3:4] > min_score
|
||||
sum_poses = np.sum(poses_3d * mask, axis=0)
|
||||
sum_mask = np.sum(mask, axis=0)
|
||||
imask = poses_3d[:, :, 3:4] > min_score
|
||||
sum_poses = np.sum(poses_3d * imask, axis=0)
|
||||
sum_mask = np.sum(imask, axis=0)
|
||||
initial_pose_3d = np.divide(
|
||||
sum_poses, sum_mask, where=(sum_mask > 0), out=np.zeros_like(sum_poses)
|
||||
)
|
||||
|
||||
# Use center as default if the initial pose is empty
|
||||
jmask = initial_pose_3d[:, 3] > 0.0
|
||||
sum_joints = np.sum(initial_pose_3d[jmask, 0:3], axis=0)
|
||||
sum_mask = np.sum(jmask)
|
||||
center = np.divide(
|
||||
sum_joints, sum_mask, where=(sum_mask > 0), out=np.zeros_like(sum_joints)
|
||||
)
|
||||
initial_pose_3d[~jmask, 0:3] = center
|
||||
|
||||
# Drop joints with low scores
|
||||
offset = 0.1
|
||||
mask = poses_3d[:, :, 3:4] > (min_score - offset)
|
||||
|
||||
# Drop outliers that are far away from the other proposals
|
||||
max_dist = 0.3
|
||||
max_dist = 1.2
|
||||
distances = np.linalg.norm(
|
||||
poses_3d[:, :, :3] - initial_pose_3d[np.newaxis, :, :3], axis=2
|
||||
)
|
||||
dist_mask = distances <= max_dist
|
||||
mask = mask & np.expand_dims(dist_mask, axis=-1)
|
||||
dmask = distances <= max_dist
|
||||
mask = mask & np.expand_dims(dmask, axis=-1)
|
||||
|
||||
# Select the best-k proposals for each joint that are closest to the initial pose
|
||||
keep_best = 3
|
||||
@ -265,13 +285,24 @@ def merge_group(poses_3d: np.ndarray):
|
||||
# ==================================================================================================
|
||||
|
||||
|
||||
def get_3d_pose(poses_2d, camparams, roomparams, joint_names_2d):
|
||||
def get_3d_pose(poses_2d, camparams, roomparams, joint_names_2d, min_score=0.95):
|
||||
"""Triangulates 3D poses from 2D poses of multiple views"""
|
||||
|
||||
# Convert poses and camparams to numpy arrays
|
||||
camparams = copy.deepcopy(camparams)
|
||||
for i in range(len(camparams)):
|
||||
poses_2d[i] = np.asarray(poses_2d[i])
|
||||
camparams[i]["K"] = np.array(camparams[i]["K"])
|
||||
camparams[i]["R"] = np.array(camparams[i]["R"])
|
||||
camparams[i]["T"] = np.array(camparams[i]["T"])
|
||||
camparams[i]["DC"] = np.array(camparams[i]["DC"][0:5])
|
||||
|
||||
# Undistort 2D points
|
||||
for i, cam in enumerate(camparams):
|
||||
poses = np.array(poses_2d[i])
|
||||
poses[:, :, 0:2], camparams[i] = undistort_points(poses[:, :, 0:2], cam)
|
||||
poses_2d[i] = poses.tolist()
|
||||
for i in range(len(camparams)):
|
||||
poses = poses_2d[i]
|
||||
cam = camparams[i]
|
||||
poses[:, :, 0:2] = undistort_points(poses[:, :, 0:2], cam)
|
||||
poses_2d[i] = poses
|
||||
|
||||
# Create pairs of persons
|
||||
num_persons = [len(p) for p in poses_2d]
|
||||
@ -302,28 +333,27 @@ def get_3d_pose(poses_2d, camparams, roomparams, joint_names_2d):
|
||||
# draw_utils.utils_view.show_plots()
|
||||
|
||||
# Drop pairs with low scores
|
||||
min_score = 0.9
|
||||
all_pairs = [p for p in all_pairs if p[2][1] > min_score]
|
||||
|
||||
# Group pairs that share a person
|
||||
groups = calc_grouping(all_pairs, min_score)
|
||||
|
||||
# Calculate full 3D poses
|
||||
poses_3d = []
|
||||
for pair in all_pairs:
|
||||
cam1 = camparams[pair[0][0]]
|
||||
cam2 = camparams[pair[0][1]]
|
||||
pose1 = np.array(poses_2d[pair[0][0]][pair[0][2]])
|
||||
pose2 = np.array(poses_2d[pair[0][1]][pair[0][3]])
|
||||
pose1 = poses_2d[pair[0][0]][pair[0][2]]
|
||||
pose2 = poses_2d[pair[0][1]][pair[0][3]]
|
||||
|
||||
pose_3d, _ = calc_pose_scored(pose1, pose2, cam1, cam2, roomparams)
|
||||
pair.append(pose_3d)
|
||||
|
||||
# Group pairs that share a person
|
||||
groups = calc_grouping(all_pairs)
|
||||
|
||||
# Merge groups
|
||||
poses_3d = []
|
||||
for group in groups:
|
||||
poses = np.array([p[3] for p in group[2]])
|
||||
pose_3d = merge_group(poses)
|
||||
poses = np.array([p[4] for p in group[2]])
|
||||
pose_3d = merge_group(poses, min_score)
|
||||
poses_3d.append(pose_3d)
|
||||
|
||||
if len(poses_3d) > 0:
|
||||
|
||||
2
skelda
2
skelda
Submodule skelda updated: dcf3a0c3df...85db8d366d
Reference in New Issue
Block a user