def main():
args = parse_args()
# prepare model (SiamRPN or SiamFC)
# prepare tracker
info = edict()
info.arch = args.arch
info.dataset = 'VOT2019'
info.TRT = 'TRT' in args.arch
info.epoch_test = False
siam_info = edict()
siam_info.arch = args.arch
siam_info.dataset = 'VOT2019'
siam_info.online = args.online
siam_info.epoch_test = False
siam_info.TRT = 'TRT' in args.arch
siam_info.align = False
if siam_info.TRT:
siam_info.align = False
siam_tracker = DAG(siam_info)
siam_net = models.__dict__[args.arch](align=siam_info.align, online=args.online)
print('===> init Siamese <====')
if not siam_info.TRT:
siam_net = load_pretrain(siam_net, args.resume)
else:
print("tensorrt toy model: not loading checkpoint")
siam_net.eval()
siam_net = siam_net.cuda()
if siam_info.TRT:
print('===> load model from TRT <===')
print('===> please ignore the warning information of TRT <===')
print('===> We only provide a toy demo for TensorRT. There are some operations are not supported well.<===')
print('===> If you wang to test on benchmark, please us Pytorch version. <===')
print('===> The tensorrt code will be contingously optimized (with the updating of official TensorRT.)<===')
trtNet = reloadTRT()
siam_net.tensorrt_init(trtNet)
if args.online:
online_tracker = ONLINE(info)
else:
online_tracker = None
print('[*] ======= Track video with {} ======='.format(args.arch))
# check init box is list or not
if not isinstance(args.init_bbox, list) and args.init_bbox is not None:
args.init_bbox = list(eval(args.init_bbox))
else:
args.init_bbox = None
print('===> please draw a box with your mouse <====')
track_video(siam_tracker, online_tracker, siam_net, args.video, init_box=args.init_bbox, args=args)
def main():
print('Warning: this is a demo to test OceanPlus')
print(
'Warning: if you want to test it on VOT2020, please use our integration scripts'
)
args = parse_args()
info = edict()
info.arch = args.arch
info.dataset = args.dataset
info.online = args.online
info.TRT = 'TRT' in args.arch
siam_info = edict()
siam_info.arch = args.arch
siam_info.dataset = args.dataset
siam_info.vis = args.vis
siam_tracker = OceanPlus(siam_info)
if args.mms == 'True':
MMS = True
else:
MMS = False
siam_net = models.__dict__[args.arch](online=args.online, mms=MMS)
print('===> init Siamese <====')
siam_net = load_pretrain(siam_net, args.resume)
siam_net.eval()
siam_net = siam_net.cuda()
# if info.TRT:
# print('===> load model from TRT <===')
# print('===> please ignore the warning information of TRT <===')
# trtNet = reloadTRT()
# siam_net.tensorrt_init(trtNet)
if args.online:
online_tracker = ONLINE(info)
else:
online_tracker = None
print('====> warm up <====')
for i in tqdm(range(20)):
siam_net.template(
torch.rand(1, 3, 127, 127).cuda(),
torch.rand(1, 127, 127).cuda())
siam_net.track(torch.rand(1, 3, 255, 255).cuda())
# prepare video
print('====> load dataset <====')
dataset = load_dataset(args.dataset)
video_keys = list(dataset.keys()).copy()
# hyper-parameters in or not
if args.hp is None:
hp = None
elif isinstance(args.hp, str):
f = open(join('tune', args.hp), 'r')
hp = json.load(f)
f.close()
print('====> tuning hp: {} <===='.format(hp))
else:
raise ValueError('not supported hyper-parameters')
# tracking all videos in benchmark
for video in video_keys:
if args.dataset in ['DAVIS2016', 'DAVIS2017', 'YTBVOS']: # VOS
track_vos(siam_tracker, online_tracker, siam_net, dataset[video],
args, hp)
else: # VOTS (i.e. VOT2020)
if video == 'butterfly':
track(siam_tracker, online_tracker, siam_net, dataset[video],
args)
def main():
args = parse_args()
info = edict()
info.arch = args.arch
info.dataset = args.dataset
info.TRT = 'TRT' in args.arch
info.epoch_test = args.epoch_test
siam_info = edict()
siam_info.arch = args.arch
siam_info.dataset = args.dataset
siam_info.online = args.online
siam_info.epoch_test = args.epoch_test
siam_info.TRT = 'TRT' in args.arch
if args.online:
siam_info.align = False
else:
siam_info.align = True if 'VOT' in args.dataset and args.align == 'True' else False
if siam_info.TRT:
siam_info.align = False
siam_tracker = Ocean(siam_info)
siam_net = models.__dict__[args.arch](align=siam_info.align,
online=args.online)
print(siam_net)
print('===> init Siamese <====')
if not siam_info.TRT:
siam_net = load_pretrain(siam_net, args.resume)
else:
print("tensorrt toy model: not loading checkpoint")
siam_net.eval()
siam_net = siam_net.cuda()
if siam_info.TRT:
print('===> load model from TRT <===')
print('===> please ignore the warning information of TRT <===')
print(
'===> We only provide a toy demo for TensorRT. There are some operations are not supported well.<==='
)
print(
'===> If you wang to test on benchmark, please us Pytorch version. <==='
)
print(
'===> The tensorrt code will be contingously optimized (with the updating of official TensorRT.)<==='
)
trtNet = reloadTRT()
siam_net.tensorrt_init(trtNet)
if args.online:
online_tracker = ONLINE(info)
else:
online_tracker = None
print('====> warm up <====')
for i in tqdm(range(10)):
siam_net.template(torch.rand(1, 3, 127, 127).cuda())
siam_net.track(torch.rand(1, 3, 255, 255).cuda())
# prepare video
dataset = load_dataset(args.dataset)
video_keys = list(dataset.keys()).copy()
if args.video is not None:
track(siam_tracker, online_tracker, siam_net, dataset[args.video],
args)
else:
for video in video_keys:
track(siam_tracker, online_tracker, siam_net, dataset[video], args)