def __init__(self, net_path=None):
super().init(name='ACTTracker', is_deterministic=True)
np.random.seed(123)
torch.manual_seed(456)
torch.cuda.manual_seed(789)
self.model = MDNet()
self.actor = Actor()
self.result = []
self.result_bb = []
self.success = 1
if opts['use_gpu']:
self.model = self.model.cuda()
self.actor = self.actor.cuda()
self.model.set_learnable_params(opts['ft_layers'])
self.criterion = BinaryLoss()
self.init_optimizer = set_optimizer(self.model, opts['lr_init'])
self.update_optimizer = set_optimizer(self.model, opts['lr_update'])
self.detetion = 0
self.frame = 0
def __init__(self):
super(RLS_RTMDNet, self).__init__("RLS-RTMDNet")
opts["model_path"] = path_config.RLS_RTMDNET_MODEL
opts["visual_log"] = False
opts["set_type"] = "OTB100"
opts["visualize"] = False
opts["adaptive_align"] = True
opts["padding"] = 1.2
opts["jitter"] = True
self.dtype = torch.cuda.FloatTensor # run on GPU
# Init model
self.model = MDNet(opts["model_path"])
if opts["adaptive_align"]:
align_h = self.model.roi_align_model.aligned_height
align_w = self.model.roi_align_model.aligned_width
spatial_s = self.model.roi_align_model.spatial_scale
self.model.roi_align_model = RoIAlignAdaMax(
align_h, align_w, spatial_s)
if opts["use_gpu"]:
self.model = self.model.cuda()
self.model.set_learnable_params(opts["ft_layers"])
# Init image crop model
self.img_crop_model = imgCropper(1.0)
if opts["use_gpu"]:
self.img_crop_model.gpuEnable()
# Init criterion and optimizer
self.criterion = BinaryLoss()
self.init_optimizer = set_optimizer(self.model, 0.01)
self.update_optimizer = set_optimizer(self.model, opts["lr_update"])
self.update_optimizer_owm = set_optimizer(self.model, 0.01)
# MAIN
handle = VOT("rectangle")
selection = handle.region()
imagefile = handle.frame()
if not imagefile:
sys.exit(0)
gt_rect = [selection.x, selection.y, selection.width, selection.height]
np.random.seed(0)
torch.manual_seed(0)
target_bbox = np.array(gt_rect)
model = MDNet(opts['model_path'])
if opts['use_gpu']:
model = model.cuda()
# Init criterion and optimizer
criterion = BCELoss()
model.set_learnable_params(opts['ft_layers'])
init_optimizer = set_optimizer(model, opts['lr_init'], opts['lr_mult'])
update_optimizer = set_optimizer(model, opts['lr_update'], opts['lr_mult'])
# Load first image
image = Image.open(imagefile).convert('RGB')
# Draw pos/neg samples
pos_examples = SampleGenerator('gaussian', image.size, opts['trans_pos'],
opts['scale_pos'])(target_bbox,
opts['n_pos_init'],
opts['overlap_pos_init'])
def train_mdnet(opts):
# Init dataset
with open(opts['data_path'], 'rb') as fp:
data = pickle.load(fp)
K = len(data)
dataset = [None] * K
for k, seq in enumerate(data.values()):
dataset[k] = RegionDataset(seq['images'], seq['gt'], opts)
# Init model
model = MDNet(opts['init_model_path'], K)
if opts['use_gpu']:
model = model.cuda()
model.set_learnable_params(opts['ft_layers'])
# Init criterion and optimizer
criterion = BCELoss()
evaluator = Precision()
optimizer = set_optimizer(model, opts['lr'], opts['lr_mult'])
# Main trainig loop
for i in range(opts['n_cycles']):
print('==== Start Cycle {:d}/{:d} ===='.format(i + 1, opts['n_cycles']))
if i in opts.get('lr_decay', []):
print('decay learning rate')
for param_group in optimizer.param_groups:
param_group['lr'] *= opts.get('gamma', 0.1)
# Training
model.train()
prec = np.zeros(K)
k_list = np.random.permutation(K)
for j, k in enumerate(k_list):
tic = time.time()
# training
pos_regions, neg_regions = dataset[k].next()
if opts['use_gpu']:
pos_regions = pos_regions.cuda()
neg_regions = neg_regions.cuda()
pos_score = model(pos_regions, k)
neg_score = model(neg_regions, k)
loss = criterion(pos_score, neg_score)
batch_accum = opts.get('batch_accum', 1)
if j % batch_accum == 0:
model.zero_grad()
loss.backward()
if j % batch_accum == batch_accum - 1 or j == len(k_list) - 1:
if 'grad_clip' in opts:
torch.nn.utils.clip_grad_norm_(model.parameters(), opts['grad_clip'])
optimizer.step()
prec[k] = evaluator(pos_score, neg_score)
toc = time.time()-tic
print('Cycle {:2d}/{:2d}, Iter {:2d}/{:2d} (Domain {:2d}), Loss {:.3f}, Precision {:.3f}, Time {:.3f}'
.format(i, opts['n_cycles'], j, len(k_list), k, loss.item(), prec[k], toc))
print('Mean Precision: {:.3f}'.format(prec.mean()))
print('Save model to {:s}'.format(opts['model_path']))
if opts['use_gpu']:
model = model.cpu()
states = {'shared_layers': model.layers.state_dict()}
torch.save(states, opts['model_path'])
if opts['use_gpu']:
model = model.cuda()
def run_mdnet(img_list, init_bbox, gt=None, savefig_dir='', display=False):
# Init bbox
target_bbox = np.array(init_bbox)
result = np.zeros((len(img_list), 4))
result_bb = np.zeros((len(img_list), 4))
result[0] = target_bbox
result_bb[0] = target_bbox
if gt is not None:
overlap = np.zeros(len(img_list))
overlap[0] = 1
# Init model
model = MDNet(opts['model_path'])
if opts['use_gpu']:
model = model.cuda()
# Init criterion and optimizer
criterion = BCELoss()
model.set_learnable_params(opts['ft_layers'])
init_optimizer = set_optimizer(model, opts['lr_init'], opts['lr_mult'])
update_optimizer = set_optimizer(model, opts['lr_update'], opts['lr_mult'])
tic = time.time()
# Load first image
image = Image.open(img_list[0]).convert('RGB')
# Draw pos/neg samples
pos_examples = SampleGenerator('gaussian', image.size, opts['trans_pos'],
opts['scale_pos'])(target_bbox,
opts['n_pos_init'],
opts['overlap_pos_init'])
neg_examples = np.concatenate([
SampleGenerator('uniform', image.size, opts['trans_neg_init'],
opts['scale_neg_init'])(target_bbox,
int(opts['n_neg_init'] * 0.5),
opts['overlap_neg_init']),
SampleGenerator('whole', image.size)(target_bbox,
int(opts['n_neg_init'] * 0.5),
opts['overlap_neg_init'])
])
neg_examples = np.random.permutation(neg_examples)
# Extract pos/neg features
pos_feats = forward_samples(model, image, pos_examples)
neg_feats = forward_samples(model, image, neg_examples)
# Initial training
train(model, criterion, init_optimizer, pos_feats, neg_feats,
opts['maxiter_init'])
del init_optimizer, neg_feats
torch.cuda.empty_cache()
# Train bbox regressor
bbreg_examples = SampleGenerator(
'uniform', image.size, opts['trans_bbreg'], opts['scale_bbreg'],
opts['aspect_bbreg'])(target_bbox, opts['n_bbreg'],
opts['overlap_bbreg'])
bbreg_feats = forward_samples(model, image, bbreg_examples)
bbreg = BBRegressor(image.size)
bbreg.train(bbreg_feats, bbreg_examples, target_bbox)
del bbreg_feats
torch.cuda.empty_cache()
# Init sample generators for update：这三个是【生成器】，不是产生的数据
sample_generator = SampleGenerator('gaussian', image.size, opts['trans'],
opts['scale'])
pos_generator = SampleGenerator('gaussian', image.size, opts['trans_pos'],
opts['scale_pos'])
neg_generator = SampleGenerator('uniform', image.size, opts['trans_neg'],
opts['scale_neg'])
# Init pos/neg features for update
neg_examples = neg_generator(target_bbox, opts['n_neg_update'],
opts['overlap_neg_init'])
neg_feats = forward_samples(model, image, neg_examples)
pos_feats_all = [pos_feats]
neg_feats_all = [neg_feats]
spf_total = time.time() - tic
# Display
savefig = savefig_dir != ''
if display or savefig:
dpi = 80.0
figsize = (image.size[0] / dpi, image.size[1] / dpi)
fig = plt.figure(frameon=False, figsize=figsize, dpi=dpi)
ax = plt.Axes(fig, [0., 0., 1., 1.])
ax.set_axis_off()
fig.add_axes(ax)
im = ax.imshow(image, aspect='auto')
if gt is not None:
gt_rect = plt.Rectangle(tuple(gt[0, :2]),
gt[0, 2],
gt[0, 3],
linewidth=3,
edgecolor="#00ff00",
zorder=1,
fill=False)
ax.add_patch(gt_rect)
rect = plt.Rectangle(tuple(result_bb[0, :2]),
result_bb[0, 2],
result_bb[0, 3],
linewidth=3,
edgecolor="#ff0000",
zorder=1,
fill=False)
ax.add_patch(rect)
if display:
plt.pause(.01)
plt.draw()
if savefig:
fig.savefig(os.path.join(savefig_dir, '0000.jpg'), dpi=dpi)
# Main loop
for i in range(1, len(img_list)):
tic = time.time()
# Load image
image = Image.open(img_list[i]).convert('RGB')
# Estimate target bbox：指的是下一次的target_bbox，用于迭代
samples = sample_generator(target_bbox, opts['n_samples'])
sample_scores = forward_samples(
model, image, samples, out_layer='fc6'
) #forward_samples是根据当前的给当前的多个sample打分数，【相当于执行了一次网络判断】
top_scores, top_idx = sample_scores[:, 1].topk(5)
top_idx = top_idx.cpu()
target_score = top_scores.mean()
target_bbox = samples[top_idx]
if top_idx.shape[0] > 1:
target_bbox = target_bbox.mean(axis=0) #多个target_bbox进行均值优化
success = target_score > 0
# Expand search area at failure：在【跟踪的同时】，根据跟踪情况，采用不同的sample生成参数
if success:
sample_generator.set_trans(opts['trans'])
else:
sample_generator.expand_trans(opts['trans_limit'])
# Bbox regression ：利用【当前：只是对当前帧图像进行回归，没有考虑前几帧】几名对应的box，最为regression的输入，来产生一个更好的Bbox
if success:
bbreg_samples = samples[top_idx]
if top_idx.shape[0] == 1:
bbreg_samples = bbreg_samples[None, :]
bbreg_feats = forward_samples(model, image, bbreg_samples)
bbreg_samples = bbreg.predict(bbreg_feats, bbreg_samples)
bbreg_bbox = bbreg_samples.mean(axis=0)
else:
bbreg_bbox = target_bbox
# Save result
result[i] = target_bbox
result_bb[i] = bbreg_bbox
# Data collect
if success:
pos_examples = pos_generator(target_bbox, opts['n_pos_update'],
opts['overlap_pos_update'])
pos_feats = forward_samples(model, image, pos_examples)
pos_feats_all.append(pos_feats)
if len(pos_feats_all) > opts['n_frames_long']:
del pos_feats_all[0]
neg_examples = neg_generator(target_bbox, opts['n_neg_update'],
opts['overlap_neg_update'])
neg_feats = forward_samples(model, image, neg_examples)
neg_feats_all.append(neg_feats)
if len(neg_feats_all) > opts['n_frames_short']:
del neg_feats_all[0]
# Short term update ：每张图片都要更新一次模型，这次的输入数据就是【前几帧累计的正样本和负样本】
if not success:
nframes = min(opts['n_frames_short'], len(pos_feats_all))
pos_data = torch.cat(pos_feats_all[-nframes:], 0)
neg_data = torch.cat(neg_feats_all, 0)
train(model, criterion, update_optimizer, pos_data, neg_data,
opts['maxiter_update'])
# Long term update
elif i % opts['long_interval'] == 0:
pos_data = torch.cat(pos_feats_all, 0)
neg_data = torch.cat(neg_feats_all, 0)
train(model, criterion, update_optimizer, pos_data, neg_data,
opts['maxiter_update'])
torch.cuda.empty_cache()
spf = time.time() - tic
spf_total += spf
# Display
if display or savefig:
im.set_data(image)
if gt is not None:
gt_rect.set_xy(gt[i, :2])
gt_rect.set_width(gt[i, 2])
gt_rect.set_height(gt[i, 3])
rect.set_xy(result_bb[i, :2])
rect.set_width(result_bb[i, 2])
rect.set_height(result_bb[i, 3])
if display:
plt.pause(.01)
plt.draw()
if savefig:
fig.savefig(os.path.join(savefig_dir, '{:04d}.jpg'.format(i)),
dpi=dpi)
if gt is None:
print('Frame {:d}/{:d}, Score {:.3f}, Time {:.3f}'.format(
i, len(img_list), target_score, spf))
else:
overlap[i] = overlap_ratio(gt[i], result_bb[i])[0]
print('Frame {:d}/{:d}, Overlap {:.3f}, Score {:.3f}, Time {:.3f}'.
format(i, len(img_list), overlap[i], target_score, spf))
if gt is not None:
print('meanIOU: {:.3f}'.format(overlap.mean()))
fps = len(img_list) / spf_total
return result, result_bb, fps
def search_track(self, track_num, frame_idx, init_bbox, previous_num,
init_conf):
# img_list, init_bbox
# Init bbox
# result = np.zeros((len(img_list), 4))
# result_bb = np.zeros((len(img_list), 4))
# result[0] = target_bbox
# result_bb[0] = target_bbox
# superorange params
track_list = [(-1, -1)] * (self.TRACKLET_NUM + 1)
bbox_list = [(frame_idx, init_bbox, init_conf)]
IOU_count = 0
# last_bbox = -1
# next_frame = -1
# frameA_path = os.path.join(self.FILE_PATH, str(frame_idx) + "."+ext)
frameA_path = os.path.join(self.FILE_PATH,
str(frame_idx).zfill(5) + ".png")
target_bbox = np.array(init_bbox)
# Init model
model = MDNet(opts['model_path'])
if opts['use_gpu']:
model = model.cuda()
# Init criterion and optimizer
criterion = BCELoss()
model.set_learnable_params(opts['ft_layers'])
init_optimizer = set_optimizer(model, opts['lr_init'], opts['lr_mult'])
update_optimizer = set_optimizer(model, opts['lr_update'],
opts['lr_mult'])
tic = time.time()
# Load first image
image = Image.open(frameA_path).convert('RGB')
# Draw pos/neg samples
pos_examples = SampleGenerator('gaussian', image.size,
opts['trans_pos'], opts['scale_pos'])(
target_bbox, opts['n_pos_init'],
opts['overlap_pos_init'])
neg_examples = np.concatenate([
SampleGenerator('uniform', image.size, opts['trans_neg_init'],
opts['scale_neg_init'])(target_bbox,
int(opts['n_neg_init'] *
0.5),
opts['overlap_neg_init']),
SampleGenerator('whole', image.size)(target_bbox,
int(opts['n_neg_init'] * 0.5),
opts['overlap_neg_init'])
])
neg_examples = np.random.permutation(neg_examples)
# Extract pos/neg features
if len(pos_examples) == 0 or len(neg_examples) == 0:
print("!!pos_examples=0 skip!!")
return
pos_feats = self.forward_samples(model, image, pos_examples)
neg_feats = self.forward_samples(model, image, neg_examples)
# Initial training
self.train(model, criterion, init_optimizer, pos_feats, neg_feats,
opts['maxiter_init'])
del init_optimizer, neg_feats
torch.cuda.empty_cache()
# Train bbox regressor
bbreg_examples = SampleGenerator(
'uniform', image.size, opts['trans_bbreg'], opts['scale_bbreg'],
opts['aspect_bbreg'])(target_bbox, opts['n_bbreg'],
opts['overlap_bbreg'])
bbreg_feats = self.forward_samples(model, image, bbreg_examples)
bbreg = BBRegressor(image.size)
bbreg.train(bbreg_feats, bbreg_examples, target_bbox)
del bbreg_feats
torch.cuda.empty_cache()
# Init sample generators for update
sample_generator = SampleGenerator('gaussian', image.size,
opts['trans'], opts['scale'])
pos_generator = SampleGenerator('gaussian', image.size,
opts['trans_pos'], opts['scale_pos'])
neg_generator = SampleGenerator('uniform', image.size,
opts['trans_neg'], opts['scale_neg'])
# Init pos/neg features for update
neg_examples = neg_generator(target_bbox, opts['n_neg_update'],
opts['overlap_neg_init'])
neg_feats = self.forward_samples(model, image, neg_examples)
pos_feats_all = [pos_feats]
neg_feats_all = [neg_feats]
spf_total = time.time() - tic
# fps = len(img_list) / spf_total
# return result, result_bb, fps
# Main loop
for i in range(0, self.TRACKLET_NUM): # next 10 frame
frameB_idx = frame_idx + i + 1
# print("frameB_idx="+str(frameB_idx))
if frameB_idx > self.frame_num:
break
else:
frameB_path = os.path.join(self.FILE_PATH,
str(frameB_idx).zfill(5) + ".png")
# ------------track by MDNet------------
# Load image
image = Image.open(frameB_path).convert('RGB')
# Estimate target bbox
samples = sample_generator(target_bbox, opts['n_samples'])
sample_scores = self.forward_samples(model,
image,
samples,
out_layer='fc6')
top_scores, top_idx = sample_scores[:, 1].topk(5)
top_idx = top_idx.cpu()
target_score = top_scores.mean()
target_bbox = samples[top_idx]
if top_idx.shape[0] > 1:
target_bbox = target_bbox.mean(axis=0)
success = target_score > 0
# Expand search area at failure
if success:
sample_generator.set_trans(opts['trans'])
else:
sample_generator.expand_trans(opts['trans_limit'])
# Bbox regression
if success:
bbreg_samples = samples[top_idx]
if top_idx.shape[0] == 1:
bbreg_samples = bbreg_samples[None, :]
bbreg_feats = self.forward_samples(model, image,
bbreg_samples)
bbreg_samples = bbreg.predict(bbreg_feats, bbreg_samples)
bbreg_bbox = bbreg_samples.mean(axis=0)
else:
bbreg_bbox = target_bbox
# Save result
# result[i] = target_bbox
# result_bb[i] = bbreg_bbox
# print(target_bbox)
# Data collect
if success:
pos_examples = pos_generator(target_bbox,
opts['n_pos_update'],
opts['overlap_pos_update'])
pos_feats = self.forward_samples(model, image,
pos_examples)
pos_feats_all.append(pos_feats)
if len(pos_feats_all) > opts['n_frames_long']:
del pos_feats_all[0]
neg_examples = neg_generator(target_bbox,
opts['n_neg_update'],
opts['overlap_neg_update'])
neg_feats = self.forward_samples(model, image,
neg_examples)
neg_feats_all.append(neg_feats)
if len(neg_feats_all) > opts['n_frames_short']:
del neg_feats_all[0]
# Short term update
if not success:
nframes = min(opts['n_frames_short'], len(pos_feats_all))
pos_data = torch.cat(pos_feats_all[-nframes:], 0)
neg_data = torch.cat(neg_feats_all, 0)
self.train(model, criterion, update_optimizer, pos_data,
neg_data, opts['maxiter_update'])
# Long term update
elif i % opts['long_interval'] == 0:
pos_data = torch.cat(pos_feats_all, 0)
neg_data = torch.cat(neg_feats_all, 0)
self.train(model, criterion, update_optimizer, pos_data,
neg_data, opts['maxiter_update'])
torch.cuda.empty_cache()
bboxT = bbreg_bbox
anyIOU = False
for bbox_id, bboxD in enumerate(frameBBoxList[frameB_idx]):
if bboxD.match == False:
IOU_ratio = overlap_ratio(bboxD.getRec(), bboxT)
print("IOUratio={}".format(IOU_ratio))
# print(IOU_ratio)
if IOU_ratio > 0.3:
# print("overlap")
track_list[i] = (frameB_idx, bbox_id)
IOU_count = IOU_count + 1
bbox_list.append(
(frameB_idx, bboxD.getRec(), bboxD.confidence))
anyIOU = True
break
if not anyIOU:
bbox_list.append((frameB_idx, bboxT, 0))
print("bbox_list=====")
print(bbox_list)
print("track_list=====")
print(track_list)
##debug show track_list
# for idx in range(0,10):
# print(track_list[idx])
# print(bbox_list[idx])
if IOU_count >= self.IOU_count_th: # add track
for idx in range(0, self.TRACKLET_NUM + 1):
bbox_id = track_list[idx][1]
if bbox_id != -1:
# print(frame_idx+idx+1)
# print(bbox_id)
frameBBoxList[frame_idx + idx + 1][bbox_id].setMatch()
next_track_frame = -1
init_bbox_next = []
bbox_length = len(bbox_list)
#!!! track_list length=10, bbox_list length=11
for ii in range(1, bbox_length):
idx = bbox_length - ii
print("{} {} {}".format(ii, idx, track_list[idx][0]))
if track_list[idx - 1][0] != -1:
next_track_frame = (-1) * ii # count from back
print("idx ={} ii={} next_track_frame={}".format(
idx, ii, next_track_frame))
init_bbox_next = bbox_list[idx]
break
# remove overlap range
start_rm = -1
while start_rm >= previous_num:
del self.GLOBAL_TRACK_LIST[track_num][-1] # rm 1 per move
# GLOBAL_TRACK_LIST[track_num].remove(-1)
start_rm = start_rm - 1
if previous_num == 0:
self.GLOBAL_TRACK_LIST.append(bbox_list)
else:
self.GLOBAL_TRACK_LIST[track_num].extend(bbox_list)
next_start_frame = frame_idx + bbox_length + next_track_frame
init_bbox_next = bbox_list[bbox_length + next_track_frame][1]
init_conf_next = bbox_list[bbox_length + next_track_frame][2]
# print(GLOBAL_TRACK_LIST)
print("( " + str(track_num) + " " + str(next_start_frame) + " " +
str(next_track_frame) + ")")
print("init_bbox_next:" + str(init_bbox_next))
self.search_track(
track_num, next_start_frame, init_bbox_next, next_track_frame,
init_conf_next
) # next_start_frame = 310 next_track_frame=-3(will be deleted)
def run_mdnet(img_list,
init_bbox,
gt=None,
savefig_dir='',
display=False,
loss_index=1,
model_path=opts['model_path'],
seq_name=None):
#def run_mdnet(k, img_list, init_bbox, gt=None, savefig_dir='', display=False,
# loss_index=1, model_path=opts['model_path'], seq_name=None):
############################
if fewer_images:
num_images = min(sequence_len_limit, len(img_list))
else:
num_images = len(img_list)
############################
# Init bbox
target_bbox = np.array(init_bbox)
result = np.zeros((len(img_list), 4))
result_bb = np.zeros((len(img_list), 4))
result[0] = target_bbox
result_bb[0] = target_bbox
# Init iou and pred_iou
iou_list = np.zeros((len(img_list), 1)) # shape: [113.1) ### list of ious
iou_list[0] = 1.0 ### in first frame gt=result_bb (by definition)
if gt is not None:
overlap = np.zeros(len(img_list))
overlap[0] = 1
# Init model
# model = MDNet(model_path=opts['model_path'],use_gpu=opts['use_gpu'])
model = MDNet(model_path=model_path, use_gpu=opts['use_gpu'])
if opts['use_gpu']:
model = model.cuda()
print('Init criterion and optimizer')
criterion = BCELoss()
model.set_learnable_params(opts['ft_layers'])
init_optimizer = set_optimizer(model, opts['lr_init'], opts['lr_mult'])
update_optimizer = set_optimizer(model, opts['lr_update'], opts['lr_mult'])
tic = time.time()
# Load first image
image = Image.open(img_list[0]).convert('RGB')
print('Draw pos/neg samples')
# Draw pos/neg samples
pos_examples = SampleGenerator('gaussian', image.size, opts['trans_pos'],
opts['scale_pos'])(target_bbox,
opts['n_pos_init'],
opts['overlap_pos_init'])
neg_examples = np.concatenate([
SampleGenerator('uniform', image.size, opts['trans_neg_init'],
opts['scale_neg_init'])(target_bbox,
int(opts['n_neg_init'] * 0.5),
opts['overlap_neg_init']),
SampleGenerator('whole', image.size)(target_bbox,
int(opts['n_neg_init'] * 0.5),
opts['overlap_neg_init'])
])
neg_examples = np.random.permutation(neg_examples)
print('Extract pos/neg features')
# Extract pos/neg features
if fewer_images: # shorter run in general, less accurate
pos_feats = forward_samples(model, image, pos_examples)
neg_feats = forward_samples(model, image, neg_examples)
else:
pos_feats = forward_samples(model, image, pos_examples)
neg_feats = forward_samples(model, image, neg_examples)
print('Initial training')
# Initial training
train(model,
criterion,
init_optimizer,
pos_feats,
neg_feats,
opts['maxiter_init'],
loss_index=loss_index) ### iou_pred_list
del init_optimizer, neg_feats
torch.cuda.empty_cache()
print('Train bbox regressor')
# Train bbox regressor
bbreg_examples = SampleGenerator(
'uniform', image.size, opts['trans_bbreg'], opts['scale_bbreg'],
opts['aspect_bbreg'])(target_bbox, opts['n_bbreg'],
opts['overlap_bbreg'])
bbreg_feats = forward_samples(
model, image,
bbreg_examples) # calc features ### shape: [927, 4608] ###
bbreg = BBRegressor(image.size)
bbreg.train(bbreg_feats, bbreg_examples, target_bbox)
del bbreg_feats
torch.cuda.empty_cache()
# Init sample generators for update
sample_generator = SampleGenerator('gaussian', image.size, opts['trans'],
opts['scale'])
pos_generator = SampleGenerator('gaussian', image.size, opts['trans_pos'],
opts['scale_pos'])
neg_generator = SampleGenerator('uniform', image.size, opts['trans_neg'],
opts['scale_neg'])
print('Init pos/neg features for update')
# Init pos/neg features for update
neg_examples = neg_generator(target_bbox, opts['n_neg_update'],
opts['overlap_neg_init'])
neg_feats = forward_samples(model, image, neg_examples)
pos_feats_all = [pos_feats]
neg_feats_all = [neg_feats]
spf_total = time.time() - tic
# Display
savefig = savefig_dir != ''
if display or savefig:
dpi = 80.0
figsize = (image.size[0] / dpi, image.size[1] / dpi)
fig = plt.figure(frameon=False, figsize=figsize, dpi=dpi)
ax = plt.Axes(fig, [0., 0., 1., 1.])
ax.set_axis_off()
fig.add_axes(ax)
im = ax.imshow(image, aspect='auto')
if gt is not None:
gt_rect = plt.Rectangle(tuple(gt[0, :2]),
gt[0, 2],
gt[0, 3],
linewidth=3,
edgecolor="#00ff00",
zorder=1,
fill=False)
ax.add_patch(gt_rect)
#################
num_gts = np.minimum(gt.shape[0], num_images)
# print('num_gts.shape: ', num_gts.shape)
gt_centers = gt[:num_gts, :2] + gt[:num_gts, 2:] / 2
result_centers = np.zeros_like(gt[:num_gts, :2])
result_centers[0] = gt_centers[0]
result_ious = np.zeros(num_gts, dtype='float64')
result_ious[0] = 1.
#################
rect = plt.Rectangle(tuple(result_bb[0, :2]),
result_bb[0, 2],
result_bb[0, 3],
linewidth=3,
edgecolor="#ff0000",
zorder=1,
fill=False)
ax.add_patch(rect)
if display:
plt.pause(.01)
plt.draw()
if savefig:
fig.savefig(os.path.join(savefig_dir, '0000.jpg'), dpi=dpi)
print('Main Loop')
# Main loop
spf_total = 0 # I don't want to take into account initialization
for i in tqdm(range(1, num_images)):
# for i in range(1, len(img_list)):
#print('Frame: ', i)
tic = time.time()
# Load image
image = Image.open(img_list[i]).convert('RGB')
#print('Estimate target bbox (in run_mdnet)')
# Estimate target bbox
samples = sample_generator(target_bbox, opts['n_samples'])
sample_scores = forward_samples(model, image, samples, out_layer='fc6')
top_scores, top_idx = sample_scores[:, 1].topk(5)
top_idx = top_idx.cpu()
target_score = top_scores.mean()
target_bbox = samples[top_idx]
if top_idx.shape[0] > 1:
target_bbox = target_bbox.mean(axis=0)
success = target_score > 0
# Expand search area at failure
if success:
sample_generator.set_trans(opts['trans'])
else:
sample_generator.expand_trans(opts['trans_limit'])
#print('Bbox regression (in run_mdnet)')
# Bbox regression
if success:
bbreg_samples = samples[top_idx]
if top_idx.shape[0] == 1:
bbreg_samples = bbreg_samples[None, :]
bbreg_feats = forward_samples(model, image, bbreg_samples)
bbreg_samples = bbreg.predict(bbreg_feats, bbreg_samples)
bbreg_bbox = bbreg_samples.mean(axis=0)
else:
bbreg_bbox = target_bbox
# Save result
result[i] = target_bbox
result_bb[i] = bbreg_bbox
iou_list[i] = overlap_ratio(gt[i], result_bb[i])
###########################################
# identify tracking failure and abort when in VOT mode
IoU = overlap_ratio(result_bb[i], gt[i])[0]
if (IoU == 0) and init_after_loss:
print(' * lost track in frame %d since init*' % (i))
result_distances = scipy.spatial.distance.cdist(
result_centers[:i], gt_centers[:i],
metric='euclidean').diagonal()
num_images_tracked = i - 1 # we don't count frame 0 and current frame (lost track)
im.set_data(image)
if gt is not None:
if i < gt.shape[0]:
gt_rect.set_xy(gt[i, :2])
gt_rect.set_width(gt[i, 2])
gt_rect.set_height(gt[i, 3])
else:
gt_rect.set_xy(np.array([np.nan, np.nan]))
gt_rect.set_width(np.nan)
gt_rect.set_height(np.nan)
rect.set_xy(result_bb[i, :2])
rect.set_width(result_bb[i, 2])
rect.set_height(result_bb[i, 3])
plt.pause(.01)
plt.draw()
print(
'Finished identify tracking failure and abort when in VOT mode'
)
return result[:
i], result_bb[:
i], num_images_tracked, spf_total, result_distances, result_ious[:
i], True
########################################
# Data collect
if success:
pos_examples = pos_generator(target_bbox, opts['n_pos_update'],
opts['overlap_pos_update'])
pos_feats = forward_samples(model, image, pos_examples)
pos_feats_all.append(pos_feats)
if len(pos_feats_all) > opts['n_frames_long']:
del pos_feats_all[0]
neg_examples = neg_generator(target_bbox, opts['n_neg_update'],
opts['overlap_neg_update'])
neg_feats = forward_samples(model, image, neg_examples)
neg_feats_all.append(neg_feats)
if len(neg_feats_all) > opts['n_frames_short']:
del neg_feats_all[0]
# Short term update
if not success:
nframes = min(opts['n_frames_short'], len(pos_feats_all))
pos_data = torch.cat(pos_feats_all[-nframes:], 0)
neg_data = torch.cat(neg_feats_all, 0)
train(model, criterion, update_optimizer, pos_data, neg_data,
opts['maxiter_update'])
# Long term update
elif i % opts['long_interval'] == 0:
pos_data = torch.cat(pos_feats_all, 0)
neg_data = torch.cat(neg_feats_all, 0)
train(model, criterion, update_optimizer, pos_data, neg_data,
opts['maxiter_update'])
torch.cuda.empty_cache()
spf = time.time() - tic
spf_total += spf
#print('Time: ', spf)
# Display
if display or savefig:
im.set_data(image)
if gt is not None:
gt_rect.set_xy(gt[i, :2])
gt_rect.set_width(gt[i, 2])
gt_rect.set_height(gt[i, 3])
#################
result_ious[i] = overlap_ratio(result_bb[i], gt[i])[0]
result_centers[i] = result_bb[i, :2] + result_bb[i, 2:] / 2
#################
rect.set_xy(result_bb[i, :2])
rect.set_width(result_bb[i, 2])
rect.set_height(result_bb[i, 3])
if display:
plt.pause(.01)
plt.draw()
if savefig:
fig.savefig(os.path.join(savefig_dir, '{:04d}.jpg'.format(i)),
dpi=dpi)
####################################
if detailed_printing:
if gt is None:
print(" Frame %d/%d, Score %.3f, Time %.3f" % \
(i, num_images-1, target_score, spf))
else:
if i < gt.shape[0]:
print(" Frame %d/%d, Overlap %.3f, Score %.3f, Time %.3f" % \
(i, num_images-1, overlap_ratio(gt[i], result_bb[i])[0], target_score, spf))
else:
print(" Frame %d/%d, Overlap %.3f, Score %.3f, Time %.3f" % \
(i, num_images-1, overlap_ratio(np.array([np.nan,np.nan,np.nan,np.nan]), result_bb[i])[0], target_score, spf))
####################################
# if gt is None:
# print('Frame {:d}/{:d}, Score {:.3f}, Time {:.3f}'
# .format(i, len(img_list), target_score, spf))
# else:
# overlap[i] = overlap_ratio(gt[i], result_bb[i])[0]
# print('Frame {:d}/{:d}, Overlap {:.3f}, Score {:.3f}, Time {:.3f}'
# .format(i, len(img_list), overlap[i], target_score, spf))
########################
plt.close()
result_distances = scipy.spatial.distance.cdist(
result_centers, gt_centers, metric='euclidean').diagonal()
num_images_tracked = num_images - 1 # I don't want to count initialization frame (i.e. frame 0)
print(' main loop finished, %d frames' % (num_images))
print('mean IoU: ', iou_list.mean())
print('Finished run_mdnet()')
return result, result_bb, num_images_tracked, spf_total, result_distances, result_ious, False
def run_vtaan(img_list, init_bbox, gt=None, savefig_dir='', display=False):
# Init bbox
target_bbox = np.array(init_bbox)
result = np.zeros((len(img_list), 4))
result_bb = np.zeros((len(img_list), 4))
result[0] = target_bbox
result_bb[0] = target_bbox
if gt is not None:
overlap = np.zeros(len(img_list))
overlap[0] = 1
# Init model
model = MDNet(opts['model_path'])
model_g = NetG()
if opts['use_gpu']:
model = model.cuda()
model_g = model_g.cuda()
GBP = guided_backprop.GuidedBackprop(model, 1)
# Init criterion and optimizer
criterion = BCELoss()
criterion_g = torch.nn.MSELoss(reduction='sum')
model.set_learnable_params(opts['ft_layers'])
model_g.set_learnable_params(opts['ft_layers'])
init_optimizer = set_optimizer(model, opts['lr_init'], opts['lr_mult'])
update_optimizer = set_optimizer(model, opts['lr_update'], opts['lr_mult'])
tic = time.time()
# Load first image
image = Image.open(img_list[0]).convert('RGB')
# Draw pos/neg samples
pos_examples = SampleGenerator('gaussian', image.size, opts['trans_pos'],
opts['scale_pos'])(target_bbox,
opts['n_pos_init'],
opts['overlap_pos_init'])
neg_examples = np.concatenate([
SampleGenerator('uniform', image.size, opts['trans_neg_init'],
opts['scale_neg_init'])(target_bbox,
int(opts['n_neg_init'] * 0.5),
opts['overlap_neg_init']),
SampleGenerator('whole', image.size)(target_bbox,
int(opts['n_neg_init'] * 0.5),
opts['overlap_neg_init'])
])
neg_examples = np.random.permutation(neg_examples)
# Extract pos/neg features
pos_feats = forward_samples(model, image, pos_examples)
neg_feats = forward_samples(model, image, neg_examples)
pos_imgids = np.array([[0]] * pos_feats.size(0))
neg_imgids = np.array([[0]] * neg_feats.size(0))
feat_dim = pos_feats.size(-1)
# Initial training
train(model, None, criterion, init_optimizer, pos_feats, neg_feats,
opts['maxiter_init'], pos_imgids, pos_examples, neg_imgids,
neg_examples, img_list, GBP)
del init_optimizer, neg_feats
torch.cuda.empty_cache()
g_pretrain(model, model_g, criterion_g, pos_feats)
torch.cuda.empty_cache()
# Train bbox regressor
bbreg_examples = SampleGenerator(
'uniform', image.size, opts['trans_bbreg'], opts['scale_bbreg'],
opts['aspect_bbreg'])(target_bbox, opts['n_bbreg'],
opts['overlap_bbreg'])
bbreg_feats = forward_samples(model, image, bbreg_examples)
bbreg = BBRegressor(image.size)
bbreg.train(bbreg_feats, bbreg_examples, target_bbox)
del bbreg_feats
torch.cuda.empty_cache()
# Init sample generators for update
sample_generator = SampleGenerator('gaussian', image.size, opts['trans'],
opts['scale'])
pos_generator = SampleGenerator('gaussian', image.size, opts['trans_pos'],
opts['scale_pos'])
neg_generator = SampleGenerator('uniform', image.size, opts['trans_neg'],
opts['scale_neg'])
# Init pos/neg features for update
neg_examples = neg_generator(target_bbox, opts['n_neg_update'],
opts['overlap_neg_init'])
neg_feats = forward_samples(model, image, neg_examples)
pos_feats_all = [pos_feats[:opts['n_pos_update']]]
neg_feats_all = [neg_feats[:opts['n_neg_update']]]
pos_examples_all = [pos_examples[:opts['n_pos_update']]]
neg_examples_all = [neg_examples[:opts['n_neg_update']]]
pos_imgids_all = [pos_imgids[:opts['n_pos_update']]]
neg_imgids_all = [neg_imgids[:opts['n_neg_update']]]
spf_total = time.time() - tic
# Display
savefig = savefig_dir != ''
if display or savefig:
dpi = 80.0
figsize = (image.size[0] / dpi, image.size[1] / dpi)
fig = plt.figure(frameon=False, figsize=figsize, dpi=dpi)
ax = plt.Axes(fig, [0., 0., 1., 1.])
ax.set_axis_off()
fig.add_axes(ax)
im = ax.imshow(image, aspect='auto')
if gt is not None:
gt_rect = plt.Rectangle(tuple(gt[0, :2]),
gt[0, 2],
gt[0, 3],
linewidth=3,
edgecolor="#00ff00",
zorder=1,
fill=False)
ax.add_patch(gt_rect)
rect = plt.Rectangle(tuple(result_bb[0, :2]),
result_bb[0, 2],
result_bb[0, 3],
linewidth=3,
edgecolor="#ff0000",
zorder=1,
fill=False)
ax.add_patch(rect)
if display:
plt.pause(.01)
plt.draw()
if savefig:
fig.savefig(os.path.join(savefig_dir, '0000.jpg'), dpi=dpi)
# Main loop
for i in range(1, len(img_list)):
tic = time.time()
# Load image
image = Image.open(img_list[i]).convert('RGB')
# Estimate target bbox
samples = sample_generator(target_bbox, opts['n_samples'])
sample_scores = forward_samples(model, image, samples, out_layer='fc6')
top_scores, top_idx = sample_scores[:, 1].topk(5)
top_idx = top_idx.cpu()
target_score = top_scores.mean()
target_bbox = samples[top_idx]
if top_idx.shape[0] > 1:
target_bbox = target_bbox.mean(axis=0)
success = target_score > 0
# Expand search area at failure
if success:
sample_generator.set_trans(opts['trans'])
else:
sample_generator.expand_trans(opts['trans_limit'])
# Bbox regression
if success:
bbreg_samples = samples[top_idx]
if top_idx.shape[0] == 1:
bbreg_samples = bbreg_samples[None, :]
bbreg_feats = forward_samples(model, image, bbreg_samples)
bbreg_samples = bbreg.predict(bbreg_feats, bbreg_samples)
bbreg_bbox = bbreg_samples.mean(axis=0)
else:
bbreg_bbox = target_bbox
# Save result
result[i] = target_bbox
result_bb[i] = bbreg_bbox
# Data collect
if success:
pos_examples = pos_generator(target_bbox, opts['n_pos_update'],
opts['overlap_pos_update'])
pos_feats = forward_samples(model, image, pos_examples)
pos_feats_all.append(pos_feats)
if len(pos_feats_all) > opts['n_frames_long']:
del pos_feats_all[0]
del pos_examples_all[0]
del pos_imgids_all[0]
neg_examples = neg_generator(target_bbox, opts['n_neg_update'],
opts['overlap_neg_update'])
neg_feats = forward_samples(model, image, neg_examples)
neg_feats_all.append(neg_feats)
pos_examples_all.append(pos_examples)
neg_examples_all.append(neg_examples)
pos_imgids_all.append(np.array([[i]] * pos_feats.size(0)))
neg_imgids_all.append(np.array([[i]] * neg_feats.size(0)))
if len(neg_feats_all) > opts['n_frames_short']:
del neg_feats_all[0]
del neg_examples_all[0]
del neg_imgids_all[0]
# Short term update
if not success:
nframes = min(opts['n_frames_short'], len(pos_feats_all))
pos_data = torch.stack(pos_feats_all[-nframes:],
0).view(-1, feat_dim)
neg_data = torch.stack(neg_feats_all, 0).view(-1, feat_dim)
pos_examples_data = torch.from_numpy(
np.stack(pos_examples_all[-nframes:], 0)).view(-1, 4).numpy()
neg_examples_data = torch.from_numpy(np.stack(neg_examples_all,
0)).view(-1,
4).numpy()
pos_imgids_data = torch.from_numpy(
np.stack(pos_imgids_all[-nframes:], 0)).view(-1, 1).numpy()
neg_imgids_data = torch.from_numpy(np.stack(neg_imgids_all,
0)).view(-1,
1).numpy()
train(model, None, criterion, update_optimizer, pos_data, neg_data,
opts['maxiter_update'], pos_imgids_data, pos_examples_data,
neg_imgids_data, neg_examples_data, img_list, GBP)
# Long term update
elif i % opts['long_interval'] == 0:
pos_data = t.stack(pos_feats_all, 0).view(-1, feat_dim)
neg_data = t.stack(neg_feats_all, 0).view(-1, feat_dim)
pos_examples_data = torch.from_numpy(np.stack(pos_examples_all,
0)).view(-1,
4).numpy()
neg_examples_data = torch.from_numpy(np.stack(neg_examples_all,
0)).view(-1,
4).numpy()
pos_imgids_data = torch.from_numpy(np.stack(pos_imgids_all,
0)).view(-1,
1).numpy()
neg_imgids_data = torch.from_numpy(np.stack(neg_imgids_all,
0)).view(-1,
1).numpy()
# train(model, model_g, criterion, update_optimizer, pos_data, neg_data, opts['maxiter_update'],
# pos_imgids_data, pos_examples_data, neg_imgids_data, neg_examples_data, img_list, GBP)
train(model, model_g, criterion, update_optimizer, pos_data,
neg_data, opts['maxiter_update'], None, None, None, None,
img_list, GBP)
torch.cuda.empty_cache()
spf = time.time() - tic
spf_total += spf
# Display
if display or savefig:
im.set_data(image)
if gt is not None:
gt_rect.set_xy(gt[i, :2])
gt_rect.set_width(gt[i, 2])
gt_rect.set_height(gt[i, 3])
rect.set_xy(result_bb[i, :2])
rect.set_width(result_bb[i, 2])
rect.set_height(result_bb[i, 3])
if display:
plt.pause(.01)
plt.draw()
if savefig:
fig.savefig(os.path.join(savefig_dir, '{:04d}.jpg'.format(i)),
dpi=dpi)
if gt is None:
print('Frame {:d}/{:d}, Score {:.3f}, Time {:.3f}'.format(
i + 1, len(img_list), target_score, spf))
else:
overlap[i] = overlap_ratio(gt[i], result_bb[i])[0]
print('Frame {:d}/{:d}, Overlap {:.3f}, Score {:.3f}, Time {:.3f}'.
format(i + 1, len(img_list), overlap[i], target_score, spf))
if gt is not None:
print('meanIOU: {:.3f}'.format(overlap.mean()))
fps = len(img_list) / spf_total
return result, result_bb, fps