def visualize_detections(cfg, img, detections):
vis_scale = 1.0
marker_size = 4
minx = 2 * marker_size
miny = 2 * marker_size
maxx = img.shape[1] - 2 * marker_size
maxy = img.shape[0] - 2 * marker_size
unPos = detections.coord
joints_to_visualise = range(cfg.num_joints)
visim_dets = img.copy()
for pidx in joints_to_visualise:
for didx in range(unPos[pidx].shape[0]):
cur_x = unPos[pidx][didx, 0] * vis_scale
cur_y = unPos[pidx][didx, 1] * vis_scale
# / cfg.global_scale
if check_point(cur_x, cur_y, minx, miny, maxx, maxy):
_npcircle(visim_dets,
cur_x, cur_y,
marker_size,
keypoint_colors[pidx])
return visim_dets
def draw(self, visim, dataset, person_conf, image):
minx = 2 * marker_size
coor = []
miny = 2 * marker_size
maxx = visim.shape[1] - 2 * marker_size
maxy = visim.shape[0] - 2 * marker_size
num_people = person_conf.shape[0]
color_assignment = dict()
# MA: assign same color to matching body configurations
if self.prev_person_conf.shape[0] > 0 and person_conf.shape[0] > 0:
ref_points = get_ref_points(person_conf)
prev_ref_points = get_ref_points(self.prev_person_conf)
# MA: this munkres implementation assumes that num(rows) >= num(columns)
if person_conf.shape[0] <= self.prev_person_conf.shape[0]:
cost_matrix = scipy.spatial.distance.cdist(
ref_points, prev_ref_points)
else:
cost_matrix = scipy.spatial.distance.cdist(
prev_ref_points, ref_points)
assert (cost_matrix.shape[0] <= cost_matrix.shape[1])
conf_assign = self.mk.compute(cost_matrix)
if person_conf.shape[0] > self.prev_person_conf.shape[0]:
conf_assign = [(idx2, idx1) for idx1, idx2 in conf_assign]
cost_matrix = cost_matrix.T
for pidx1, pidx2 in conf_assign:
if cost_matrix[pidx1][pidx2] < min_match_dist:
color_assignment[pidx1] = self.prev_color_assignment[pidx2]
print("#tracked objects:", len(color_assignment))
free_coloridx = sorted(list(
set(range(len(self.track_colors))).difference(
set(color_assignment.values()))),
reverse=True)
for pidx in range(num_people):
# color_idx = pidx % len(self.track_colors)
if pidx in color_assignment:
color_idx = color_assignment[pidx]
else:
if len(free_coloridx) > 0:
color_idx = free_coloridx[-1]
free_coloridx = free_coloridx[:-1]
else:
color_idx = np.random.randint(len(self.track_colors))
color_assignment[pidx] = color_idx
assert (color_idx < len(self.track_colors))
if np.sum(person_conf[pidx, :, 0] > 0) < draw_conf_min_count:
continue
for kidx1, kidx2 in dataset.get_pose_segments():
p1 = (int(math.floor(person_conf[pidx, kidx1, 0])),
int(math.floor(person_conf[pidx, kidx1, 1])))
p2 = (int(math.floor(person_conf[pidx, kidx2, 0])),
int(math.floor(person_conf[pidx, kidx2, 1])))
if check_point(p1[0], p1[1], minx, miny,
maxx, maxy) and check_point(
p2[0], p2[1], minx, miny, maxx, maxy):
color = np.array(self.track_colors[color_idx][::-1],
dtype=np.float64) / 255.0
#plt.plot([p1[0], p2[0]], [p1[1], p2[1]], marker='o', linestyle='solid', linewidth=2.0, color=color)
cv2.line(image, (p1[0], p1[1]), (p2[0], p2[1]),
(0, 255, 0), 2)
#print(p1, p2 )
coor.append([p1, p2])
self.prev_person_conf = person_conf
self.prev_color_assignment = color_assignment
return coor
def draw(self, visim, dataset, person_conf):
minx = 2 * marker_size
miny = 2 * marker_size
maxx = visim.shape[1] - 2 * marker_size
maxy = visim.shape[0] - 2 * marker_size
num_people = person_conf.shape[0]
color_assignment = dict()
# MA: assign same color to matching body configurations
if self.prev_person_conf.shape[0] > 0 and person_conf.shape[0] > 0:
ref_points = get_ref_points(person_conf)
prev_ref_points = get_ref_points(self.prev_person_conf)
# MA: this munkres implementation assumes that num(rows) >= num(columns)
if person_conf.shape[0] <= self.prev_person_conf.shape[0]:
cost_matrix = scipy.spatial.distance.cdist(
ref_points, prev_ref_points)
else:
cost_matrix = scipy.spatial.distance.cdist(
prev_ref_points, ref_points)
assert (cost_matrix.shape[0] <= cost_matrix.shape[1])
conf_assign = self.mk.compute(cost_matrix)
if person_conf.shape[0] > self.prev_person_conf.shape[0]:
conf_assign = [(idx2, idx1) for idx1, idx2 in conf_assign]
cost_matrix = cost_matrix.T
for pidx1, pidx2 in conf_assign:
if cost_matrix[pidx1][pidx2] < min_match_dist:
color_assignment[pidx1] = self.prev_color_assignment[pidx2]
print("#tracked objects:", len(color_assignment))
free_coloridx = sorted(list(
set(range(len(self.track_colors))).difference(
set(color_assignment.values()))),
reverse=True)
for pidx in range(num_people):
# color_idx = pidx % len(self.track_colors)
if pidx in color_assignment:
color_idx = color_assignment[pidx]
else:
if len(free_coloridx) > 0:
color_idx = free_coloridx[-1]
free_coloridx = free_coloridx[:-1]
else:
color_idx = np.random.randint(len(self.track_colors))
color_assignment[pidx] = color_idx
assert (color_idx < len(self.track_colors))
if np.sum(person_conf[pidx, :, 0] > 0) < draw_conf_min_count:
continue
for kidx1, kidx2 in dataset.get_pose_segments():
p1 = (int(math.floor(person_conf[pidx, kidx1, 0])),
int(math.floor(person_conf[pidx, kidx1, 1])))
p2 = (int(math.floor(person_conf[pidx, kidx2, 0])),
int(math.floor(person_conf[pidx, kidx2, 1])))
if check_point(p1[0], p1[1], minx, miny,
maxx, maxy) and check_point(
p2[0], p2[1], minx, miny, maxx, maxy):
# cv2.line(visim, p1, p2, self.track_colors[color_idx][::-1], 3, cv2.CV_AA)
rr, cc, val = draw.line_aa(p1[1], p1[0], p2[1], p2[0])
color = np.array(self.track_colors[color_idx][::-1])
val = np.reshape(val, (-1, 1))
val = np.repeat(val, 3, axis=1)
visim[rr, cc, :] = val * color
for kidx in range(dataset.num_keypoints()):
cur_x = person_conf[pidx][kidx, 0]
cur_y = person_conf[pidx][kidx, 1]
if check_point(cur_x, cur_y, minx, miny, maxx, maxy):
_npcircle(visim, cur_x, cur_y, marker_size,
self.track_colors[color_idx][::-1], 0.2)
self.prev_person_conf = person_conf
self.prev_color_assignment = color_assignment
