def init(self, frame, bbox):
self.size = (frame.shape[1], frame.shape[0])
# for update
self.pos_generator = SampleGenerator('gaussian', self.size, opts['trans_pos'], opts['scale_pos'])
self.neg_generator = SampleGenerator('uniform', self.size, opts['trans_neg'], opts['scale_neg'])
# generate bounding box rect
pos_rects = gen_bboxes(SampleGenerator('gaussian', self.size, opts['trans_pos'], opts['scale_pos']),
bbox, opts['n_pos_init'], opts['overlap_pos_init'])
overlap_pos = opts['overlap_pos_init'].copy()
while len(pos_rects) == 0 and overlap_pos[0] > 0.3:
overlap_pos[0] -= 0.1
pos_rects = gen_bboxes(SampleGenerator('gaussian', self.size, opts['trans_pos'], opts['scale_pos']),
bbox, opts['n_pos_init'], overlap_pos)
neg_rects = np.random.permutation(np.concatenate([
gen_bboxes(SampleGenerator('uniform', self.size, opts['trans_neg_init'], opts['scale_neg_init']),
bbox, opts['n_neg_init'] // 2, opts['overlap_neg_init']),
gen_bboxes(SampleGenerator('whole', self.size),
bbox, opts['n_neg_init'] // 2, opts['overlap_neg_init'])]))
neg_rects = np.random.permutation(neg_rects)
# Extract pos/neg features
pos_feats = self.extract_feats(frame, pos_rects)
neg_feats = self.extract_feats(frame, neg_rects)
# generate corrsponding features of these rects
self.pos_manager.insert(pos_feats, overlap_ratio(pos_rects, bbox))
self.neg_manger.insert(neg_feats, overlap_ratio(neg_rects, bbox))
self.frames += 1
def __call__(self, bbox, n, overlap_range=None, scale_range=None):
if overlap_range is None and scale_range is None:
return self._gen_samples(bbox, n)
else:
samples = None
remain = n
factor = 2
while remain > 0 and factor < 16:
samples_ = self._gen_samples(bbox, remain * factor)
idx = np.ones(len(samples_), dtype=bool)
if overlap_range is not None:
r = overlap_ratio(samples_, bbox)
idx *= (r >= overlap_range[0]) * (r <= overlap_range[1])
if scale_range is not None:
s = np.prod(samples_[:, 2:], axis=1) / np.prod(bbox[2:])
idx *= (s >= scale_range[0]) * (s <= scale_range[1])
samples_ = samples_[idx, :]
samples_ = samples_[:min(remain, len(samples_))]
if samples is None:
samples = samples_
else:
samples = np.concatenate([samples, samples_])
remain = n - len(samples)
factor = factor * 2
return samples
def add_features(self, frame, rect):
# frame = np.asarray(frame)
pos_rects = gen_bboxes(self.pos_generator, rect, opts['n_pos_update'], opts['overlap_pos_update']) # 200
overlap_pos = opts['overlap_pos_update'].copy()
while len(pos_rects) == 0 and overlap_pos[0] > 0.3:
overlap_pos[0] -= 0.1
pos_rects = gen_bboxes(self.pos_generator, rect, opts['n_pos_update'], overlap_pos)
neg_rects = np.random.permutation(np.concatenate([
gen_bboxes(self.neg_generator, rect, opts['n_neg_update'] - opts['n_neg_update'] // 6, opts['overlap_neg_update']),
gen_bboxes(SampleGenerator('whole', self.size),rect, opts['n_neg_update'] // 6, [0, 0])
]))
# Extract pos/neg features
pos_feats = self.extract_feats(frame, pos_rects)
neg_feats = self.extract_feats(frame, neg_rects)
self.pos_manager.insert(pos_feats, overlap_ratio(pos_rects, rect))
self.neg_manger.insert(neg_feats, overlap_ratio(neg_rects, rect))
self.frames += 1
tracker.init()
all_iou = []
result_bb = []
spf_total = time.time() - tic
for i in range(1, len(img_list)):
tic = time.time()
try:
frame = cv2.cvtColor(cv2.imread(img_list[i]), cv2.COLOR_BGR2RGB)
except:
break
rect, bbreg_rect, score = tracker.update(frame=frame)
result_bb.append(bbreg_rect)
torch.cuda.empty_cache()
iou = overlap_ratio(bbreg_rect, gt[i])[0]
all_iou.append(iou)
print("frame: {}/{}, score: {:.3f} iou: {:.3f}".format(
i,
len(img_list) - 1, score, iou.squeeze()))
if opts['vis']:
img = cv2.cvtColor(np.array(frame), cv2.COLOR_RGB2BGR)
img = draw_img(img, bbreg_rect, gt[i], idx=i)
cv2.imshow('tracker', img)
cv2.waitKey(1)
cv2.imwrite('../result/{}/{}.'.format('Bird1',
str(i).zfill(4)),
draw_img(img, bbreg_rect, gt[i], idx=i))