/SiamDW-PyTorch

A Pytorch Implementation of 'Deeper and Wider Siamese Networks for Real-Time Visual Tracking (CVPR2019)'

Primary LanguagePython

SiamDW-PyTorch

This is a Pytorch implementation of SiamDW with train codes, which is mainly based on deeper_wider_siamese_trackers and Siamese-RPN. I re-format my code with reference to the author's official code SiamDW.

For more details about the CIR tracker please refer to the paper: Deeper and Wider Siamese Networks for Real-Time Visual Tracking by Zhipeng Zhang and Houwen Peng.

NOTE: The author proposed CIR/CIR-D unit into both SiamFC and SiamRPN, repectively denoted as SiamFC+ and SiamRPN+. Currently this repo only contained that of SiamFC+ and SiamRPN+ with backbone ResNet22 and others will be listed into future work.

The repo is still under development.

Requirements

  • python == 3.6
  • pytorch == 0.3.1
  • numpy == 1.12.1
  • opencv == 3.1.0

Training

  • data preparation

    1. Follow the instructions in Siamese-RPN to curate the VID and YTB dataset. If you also want to use GOT10K, follow the same instructions as above or download the curated data by author.

    2. Create the soft links data_curated and data_curated.lmdb to folder dataset.

  • download pretrained model

    1. Download pretrained model from OneDrive, GoogleDrive or BaiduDrive. Extracted code for BaiduDrive is 7rfu.

    2. Put them to models/pretrain directory.

  • chosse the training dataset by set the parameters in lib/utils/config.py.

    For example, if you would like to use both VID and YTB dataset to train SiamRPN+, then just simply set both VID_used and YTB_used into True.

  • choose the model to be trained by modifying train.sh, e.g, to train SiamFC+ using command

    CUDA_VISIBLE_DEVICES=0 python bin/train_siamfc.py --arch SiamFC_Res22 --resume ./models/SiamFC_Res22_mine.pth

    or to train SiamRPN+ by

    CUDA_VISIBLE_DEVICES=0 python bin/train_siamrpn.py --arch SiamRPN_Res22 --resume ./models/SiamRPN_Res22_mine.pth

Tracking

  • data preparation

    1. Create the soft link OTB2015 to folder dataset

  • start tracking by modifying test.sh as above

Benchmark result

  • OTB2015

Reference

[1] Zhipeng Zhang, Houwen Peng. Deeper and Wider Siamese Networks for Real-Time Visual Tracking. Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2019.