def __init__(self, im, init_rect, config=TrackerConfig(), gpu=True):
self.gpu = gpu
self.config = config
self.net = DCFNet(config)
self.net.load_param(config.feature_path)
self.net.eval()
if gpu:
self.net.cuda()
# confine results
target_pos, target_sz = rect1_2_cxy_wh(init_rect)
self.min_sz = np.maximum(config.min_scale_factor * target_sz, 4)
self.max_sz = np.minimum(im.shape[:2],
config.max_scale_factor * target_sz)
# crop template
window_sz = target_sz * (1 + config.padding)
bbox = cxy_wh_2_bbox(target_pos, window_sz)
patch = crop_chw(im, bbox, self.config.crop_sz)
target = patch - config.net_average_image
self.net.update(torch.Tensor(np.expand_dims(target, axis=0)).cuda())
self.target_pos, self.target_sz = target_pos, target_sz
self.patch_crop = np.zeros(
(config.num_scale, patch.shape[0], patch.shape[1], patch.shape[2]),
np.float32) # buff
class DCFNetTraker(object):
def __init__(self, im, init_rect, config=TrackerConfig(), gpu=False):
self.gpu = gpu
self.config = config
self.net = DCFNet(config)
self.net.load_param(config.feature_path)
self.net.eval()
if gpu:
self.net.cuda()
# confine results
target_pos, target_sz = rect1_2_cxy_wh(init_rect)
self.min_sz = np.maximum(config.min_scale_factor * target_sz, 4)
self.max_sz = np.minimum(im.shape[:2], config.max_scale_factor * target_sz)
# crop template
window_sz = target_sz * (1 + config.padding)
bbox = cxy_wh_2_bbox(target_pos, window_sz)
patch = crop_chw(im, bbox, self.config.crop_sz)
target = patch - config.net_average_image
self.net.update(torch.Tensor(np.expand_dims(target, axis=0)).cuda())
self.target_pos, self.target_sz = target_pos, target_sz
self.patch_crop = np.zeros((config.num_scale, patch.shape[0], patch.shape[1], patch.shape[2]), np.float32) # buff
def track(self, im):
for i in range(self.config.num_scale): # crop multi-scale search region
window_sz = self.target_sz * (self.config.scale_factor[i] * (1 + self.config.padding))
bbox = cxy_wh_2_bbox(self.target_pos, window_sz)
self.patch_crop[i, :] = crop_chw(im, bbox, self.config.crop_sz)
search = self.patch_crop - self.config.net_average_image
if self.gpu:
response = self.net(torch.Tensor(search).cuda())
else:
response = self.net(torch.Tensor(search))
peak, idx = torch.max(response.view(self.config.num_scale, -1), 1)
peak = peak.data.cpu().numpy() * self.config.scale_penalties
best_scale = np.argmax(peak)
r_max, c_max = np.unravel_index(idx[best_scale], self.config.net_input_size)
if r_max > self.config.net_input_size[0] / 2:
r_max = r_max - self.config.net_input_size[0]
if c_max > self.config.net_input_size[1] / 2:
c_max = c_max - self.config.net_input_size[1]
window_sz = self.target_sz * (self.config.scale_factor[best_scale] * (1 + self.config.padding))
self.target_pos = self.target_pos + np.array([c_max, r_max]) * window_sz / self.config.net_input_size
self.target_sz = np.minimum(np.maximum(window_sz / (1 + self.config.padding), self.min_sz), self.max_sz)
# model update
window_sz = self.target_sz * (1 + self.config.padding)
bbox = cxy_wh_2_bbox(self.target_pos, window_sz)
patch = crop_chw(im, bbox, self.config.crop_sz)
target = patch - self.config.net_average_image
self.net.update(torch.Tensor(np.expand_dims(target, axis=0)).cuda(), lr=self.config.interp_factor)
return cxy_wh_2_rect1(self.target_pos, self.target_sz) # 1-index
help='tune on which dataset')
parser.add_argument('--model', metavar='PATH', default='param.pth')
args = parser.parse_args()
dataset = args.dataset
base_path = join('dataset', dataset)
json_path = join('dataset', dataset + '.json')
annos = json.load(open(json_path, 'r'))
videos = sorted(annos.keys())
use_gpu = True
visualization = False
# default parameter and load feature extractor network
config = TrackerConfig()
net = DCFNet(config)
net.load_param(args.model)
net.eval().cuda()
speed = []
# loop videos
for video_id, video in enumerate(videos): # run without resetting
video_path_name = annos[video]['name']
init_rect = np.array(annos[video]['init_rect']).astype(np.float)
image_files = [
join(base_path, video_path_name, 'img', im_f)
for im_f in annos[video]['image_files']
]
n_images = len(image_files)
tic = time.time() # time start
output_sz = 121
lambda0 = 1e-4
padding = 2.0
output_sigma_factor = 0.1
output_sigma = crop_sz / (1 + padding) * output_sigma_factor
y = gaussian_shaped_labels(output_sigma, [output_sz, output_sz])
yf = torch.rfft(torch.Tensor(y).view(1, 1, output_sz, output_sz).cuda(),
signal_ndim=2)
# cos_window = torch.Tensor(np.outer(np.hanning(crop_sz), np.hanning(crop_sz))).cuda() # train without cos window
config = TrackerConfig()
model = DCFNet(config=config)
model.cuda()
gpu_num = torch.cuda.device_count()
print('GPU NUM: {:2d}'.format(gpu_num))
if gpu_num > 1:
model = torch.nn.DataParallel(model, list(range(gpu_num))).cuda()
criterion = nn.MSELoss(size_average=False).cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
target = torch.Tensor(config.y).cuda().unsqueeze(0).unsqueeze(0).repeat(
args.batch_size * gpu_num, 1, 1, 1) # for training