DRNet

🏆News: DRNet get the best EAO on the public benchmark dataset of the VOT2019 challenge

Introduction

The proposed tracker named DRNet, which consists of target-location network and scale regression network. They share the same backbone feature extraction network. The target-location network use a discriminative correlation filters (DCF) module to regress the confidence map. Then the response peak position of the confidence map and the target size in the previous frame is selected as the size of a dynamic anchor. The siamese scale regression network extract search and template features according to the dynamic anchor and regress the offsets to predict the final box. In particular, we introduce a novel loss called Iterative Reweighted (IR) loss to online update the DCF module and offline train the scale regression network. Specially, the best result uses the backbone feature (ResNet50 and SE_ResNet50). We use the pytracking as our framework.

Dependencies

• python 3.7.3
• pytorch (0.4.1)
• opencv (4.1.0.24)
• pytracking
• torchvison (0.2.2)
• Cuda

Preparation

1. Get the pretrained models from here, which consists of Res50.pth and SE_Res50.pth. Please put the models to DRNet/pytracking/networks/.
2. Install dependencies. bash install.sh conda_install_path pytracking

Note: We follow the installation instructions from pytracking@pytorch041. Detailed installation instructions refers to here.

Usage

1. conda activate pytracking

2. cd pytracking

3. export PYTHONPATH=$code_path:$PYTHONPATH

4. set the absolute CODE_PATH and TRAX_BUILD_PARH in DRNet/pytracking/VOT2019/tracker_DRNet.m, the example is DRNet/pytracking/VOT2019/tracker_DRNet_example.m.

5. run the vot-toolkit.