def vis_interp(movie, header, flow, bboxes, baseline=False, worst=False):
if baseline and worst:
raise "rendering mode could not be duplicated"
if baseline:
cap, out = open_video(movie, postfix="interp_w_baseline")
elif worst:
cap, out = open_video(movie, postfix="interp_w_worst")
else:
cap, out = open_video(movie, postfix="interp")
count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
pos = 0
for i in trange(count):
ret, frame = cap.read()
if ret is False or i > bboxes.size:
break
if baseline:
pos = i
frame = draw_i_frame(frame, flow[i], bboxes[pos])
elif header["pict_type"][i] == "I":
pos = i
frame = draw_i_frame(frame, flow[i], bboxes[pos])
elif worst:
frame = draw_i_frame(frame, flow[i], bboxes[pos])
else:
# bboxes[pos] is updated by reference
frame = draw_p_frame(frame, flow[i], bboxes[pos])
cv2.rectangle(frame, (width - 220, 20), (width - 20, 60), (0, 0, 0),
-1)
cv2.putText(frame, f"pict_type: {header['pict_type'][i]}",
(width - 210, 50), cv2.FONT_HERSHEY_DUPLEX, 1,
(255, 255, 255), 1)
out.write(frame)
cap.release()
out.release()
def vis_flow(movie, flow, draw=draw_flow):
cap, out = open_video(movie, postfix='flow')
count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
for i in tqdm.trange(count):
ret, frame = cap.read()
if ret is False:
break
frame = draw(frame, flow[i])
out.write(frame)
cap.release()
out.release()
def vis_histogram(movie, header, flow, bboxes):
graph_prefix = os.path.basename(movie)
if not os.path.exists(os.path.join("graph", graph_prefix)):
os.makedirs(os.path.join("graph", graph_prefix))
cap = open_video(movie, display=False)
count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
for i in trange(count):
ret, frame = cap.read()
if ret is False or i > bboxes.size:
break
draw_histogram(i, frame, flow[i], bboxes[i], prefix=graph_prefix)
cap.release()
def vis_annotate(movie, header, draw, onlyi=False):
cap, out = open_video(movie, postfix='annotate')
count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
pos = 0
for i in tqdm.trange(count):
ret, frame = cap.read()
if ret is False:
break
cv2.putText(frame, f"pict_type: {header['pict_type'][i]}",
(10, height - 10), cv2.FONT_HERSHEY_DUPLEX, 1,
(255, 255, 255), 1)
if onlyi:
if header["pict_type"][i] == "I":
pos = i
else:
pos = i
frame = draw(frame, i, pos)
out.write(frame)
cap.release()
out.release()
def vis_composed(movie, header, flow, bboxes, baseline=False, worst=False):
if baseline and worst:
raise "rendering mode could not be duplicated"
if baseline:
cap, out = open_video(movie, postfix="kalman_w_base")
elif worst:
cap, out = open_video(movie, postfix="kalman_w_worst")
else:
cap, out = open_video(movie, postfix="kalman")
count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
kalman_0 = KalmanInterpolator()
kalman_clos_0 = lambda bboxes, flow, frame: \
interp_kalman(bboxes, flow, frame, kalman_0, calc=calc_flow_mean)
kalman_1 = KalmanInterpolator()
kalman_clos_1 = lambda bboxes, flow, frame: \
interp_kalman(bboxes, flow, frame, kalman_1, calc=calc_flow_mean_grad)
color = (255, 0, 255)
pos = 0
for i in trange(count):
ret, frame = cap.read()
if ret is False or i > bboxes.size:
break
# linear
if baseline:
pos = i
bboxes_0 = bboxes[pos].copy()
bboxes_1 = bboxes[pos].copy()
frame = draw_i_frame(frame, flow[i], bboxes_0)
frame = draw_i_frame(frame, flow[i], bboxes_1, color=color)
elif header["pict_type"][i] == "I":
pos = i
bboxes_0 = bboxes[pos].copy()
bboxes_1 = bboxes[pos].copy()
frame = draw_i_frame(frame, flow[i], bboxes_0)
frame = draw_i_frame(frame, flow[i], bboxes_1, color=color)
kalman_0.reset(bboxes[pos])
kalman_1.reset(bboxes[pos])
elif worst:
frame = draw_i_frame(frame, flow[i], bboxes_0)
frame = draw_i_frame(frame, flow[i], bboxes_1, color=color)
else:
# bboxes_0 is updated by reference
frame = draw_p_frame(frame, flow[i], bboxes_0,
interp=kalman_clos_0)
frame = draw_p_frame(frame, flow[i], bboxes_1, color=color,
interp=kalman_clos_1)
cv2.rectangle(frame, (width-220, 20), (width-20, 60), (0, 0, 0), -1)
cv2.putText(frame,
f"pict_type: {header['pict_type'][i]}", (width-210, 50),
cv2.FONT_HERSHEY_DUPLEX, 1, (255, 255, 255), 1)
out.write(frame)
cap.release()
out.release()
def eval_mot16(src_id,
prefix="MOT16/train",
MOT16=MOT16,
thresh=0.0,
baseline=False,
worst=False,
display=False,
gop=12):
mot = MOT16(src_id)
bboxes = mot.pick_bboxes()
movie = join(prefix, src_id)
flow, header = get_flow(movie, prefix=".", gop=gop)
if display:
cap, out = open_video(movie, display=True)
else:
cap = open_video(movie, display=False)
count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
kalman = KalmanInterpolator()
interp_kalman_clos = lambda bboxes, flow, frame: \
interp_kalman(bboxes, flow, frame, kalman)
for index, bbox in enumerate(bboxes):
if not bbox.empty:
bboxes[index] = bbox.query(f"prob >= {thresh}")
start = time.perf_counter()
for i in range(count):
ret, frame = cap.read()
if ret is False or i > bboxes.size:
break
if baseline:
bbox = bboxes[i].copy()
if display:
frame = draw_i_frame(frame, flow[i], bbox)
mot.eval_frame(bbox)
kalman.reset(bbox)
elif header["pict_type"][i] == "I":
bbox = bboxes[i].copy()
if display:
frame = draw_i_frame(frame, flow[i], bbox)
mot.eval_frame(bbox)
kalman.reset(bbox)
elif worst:
if display:
frame = draw_i_frame(frame, flow[i], bbox)
mot.eval_frame(bbox)
else:
# bbox is updated by reference
if display:
frame = draw_p_frame(frame,
flow[i],
bbox,
interp=interp_kalman_clos)
else:
interp_kalman_clos(bbox, flow[i], frame)
mot.eval_frame(bbox)
if display:
cv2.rectangle(frame, (width - 220, 20), (width - 20, 60),
(0, 0, 0), -1)
cv2.putText(frame, f"pict_type: {header['pict_type'][i]}",
(width - 210, 50), cv2.FONT_HERSHEY_DUPLEX, 1,
(255, 255, 255), 1)
out.write(frame)
end = time.perf_counter()
print(f"{src_id}: {count/(end-start):3.1f} FPS")
# print(f"{end-start:5.2f}")
cap.release()
if display:
out.release()
def eval_mot16_pred(src_id,
model,
prefix="MOT16/train",
MOT16=MOT16,
thresh=0.0,
baseline=False,
worst=False,
display=False):
mot = MOT16(src_id)
movie = join(prefix, args.src_id)
flow, header = get_flow(movie, prefix=".")
if display:
cap, out = open_video(movie, display=True)
else:
cap = open_video(movie, display=False)
count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
start = time.perf_counter()
for i in trange(count):
ret, frame = cap.read()
if ret is False:
break
if baseline:
bbox = predict(model, frame, thresh=thresh)
if display:
frame = draw_i_frame(frame, flow[i], bbox)
mot.eval_frame(bbox)
elif header["pict_type"][i] == "I":
bbox = predict(model, frame, thresh=thresh)
if display:
frame = draw_i_frame(frame, flow[i], bbox)
mot.eval_frame(bbox)
elif worst:
if display:
frame = draw_i_frame(frame, flow[i], bbox)
mot.eval_frame(bbox)
else:
# bbox is updated by reference
if display:
frame = draw_p_frame(frame, flow[i], bbox)
else:
interp_linear(bbox, flow[i], frame)
mot.eval_frame(bbox)
if display:
cv2.rectangle(frame, (width - 220, 20), (width - 20, 60),
(0, 0, 0), -1)
cv2.putText(frame, f"pict_type: {header['pict_type'][i]}",
(width - 210, 50), cv2.FONT_HERSHEY_DUPLEX, 1,
(255, 255, 255), 1)
out.write(frame)
end = time.perf_counter()
print(f"{src_id}: {count/(end-start):3.1f} FPS")
cap.release()
if display:
out.release()