def compute_track(video):
track = []
proposals = []
object_detected = False
kalman_filter = None
for t, im in enumerate(video):
scores, bboxes = im_detect(sess, net, im)
bboxes, scores = filter_proposals(scores,
bboxes,
nms_thresh=NMS_THRESH,
conf_thresh=CONF_THRESH)
affinities = np.zeros(scores.shape)
if not object_detected:
if bboxes.shape[0] == 0:
track.append(np.zeros(6))
proposals.append(np.zeros((1, 6)))
continue
else:
object_detected = True
i = np.argmax(scores)
xhat_0 = bboxes[i]
kalman_filter = KalmanFilter(xhat_0)
track.append(
np.concatenate((bboxes[i], scores[i], affinities[i])))
proposals.append(
np.concatenate([bboxes, scores, affinities], axis=1))
continue
xhat, P = kalman_filter.update_time()
score, affinity = np.zeros(1), np.zeros(1)
if bboxes.shape[0] > 0:
track_im = crop_and_resize(im, track[-1][:4])
bbox_ims = [crop_and_resize(im, bbox) for bbox in bboxes]
affinities = [
get_affinity(track_im, bbox_im) for bbox_im in bbox_ims
]
affinities = np.reshape(affinities, scores.shape)
i, aff = np.argmax(affinities), np.max(affinities)
if aff > AFF_THRESH:
xhat, P = kalman_filter.update_measurement(bboxes[i][:4])
score, affinity = scores[i], affinities[i]
track.append(np.concatenate((xhat, score, affinity)))
proposals.append(np.concatenate((bboxes, scores, affinities), axis=1))
return track, proposals
