zj5559 / MetricNet

Online Filtering Training Samples for Robust Visual Tracking (ACM MM2020)

MetricNet

Online Filtering Training Samples for Robust Visual Tracking (ACM MM2020)

Paper link

• Baidu Yun 提取码：pitf
• Google Drive

Results

OTB2015	Success	Precision
MDNet	0.671	0.904
MDNet+MetricNet	0.681	0.910
ECO	0.666	0.903
ECO+MetricNet	0.678	0.926
ATOM	0.665	0.870
ATOM+MetricNet	0.675	0.881

UAV123	Success	Precision
MDNet	0.540	0.754
MDNet+MetricNet	0.561	0.789
ECO	0.533	0.764
ECO+MetricNet	0.546	0.786
ATOM	0.621	0.832
ATOM+MetricNet	0.650	0.866

LaSOT	Success	Norm Precision
MDNet	0.390	0.430
MDNet+MetricNet	0.443	0.523
ECO	0.371	0.431
ECO+MetricNet	0.419	0.501
ATOM	0.503	0.574
ATOM+MetricNet	0.535	0.614

Requirments

python 3.7
pytorch
ubuntu 16.04 + cuda-9.0

Installation

The pretrained models are also downloaded.

bash install.sh conda_install_path metricnet

Train

Prepare dataset (LaSOT)

cd Train
python prepare_data.py

Train MetricNet

python train.py

Eval

Integrate MetricNet into MDNet

cd MDNet_MetricNet
python metric_tracking.py

Integrate MetricNet into ECO/ATOM

cd pytracking_MetricNet/pytracking
python run_tracker.py