UPDATES:

• [2020.03.02] Update training scripts to match the settings in the paper (12 epochs on COCO and another 12 epochs on COCO + GOT + LaSOT)!
• [2020.02.19] Both training and evaluation code are available!
  
• [2020.02.19] Initial and pretrained weights are provided!
  
• [2020.02.19] A demo tracking video of GlobalTrack is available here!

Official implementation of our AAAI2020 paper: GlobalTrack: A Simple and Strong Baseline for Long-term Tracking. The first tracker with NO cumulative errors.

Extremely simple tracking process, with NO motion model, NO online learning, NO punishment on position or scale changes, NO scale smoothing and NO trajectory refinement.

Outperforms SPLT (ICCV19), SiamRPN, ATOM and MBMD on TLP benchmark (avg. 13,529 frames per video) by MORE THAN 11% (absolute gain).

Outperforms SPLT, SiamRPN++, ATOM and DaSiamLT on LaSOT benchmark.

Paper on arXiv: 1912.08531.

Demo video: YouTube, YouKu.

To reproduce our Python environment, you'll need to create a conda environment from environment.yml and compile the Cpp/CUDA extensions (we use CUDA toolkit 9.0):

conda env create -f environment.yml
conda activate GlobalTrack
git clone https://github.com/huanglianghua/GlobalTrack.git
cd _submodules/mmdetection
python setup.py develop

Alternatively, you can also install PyTorch==1.1.0, torchvision, shapely and scipy manually, then compile the Cpp/CUDA extensions by running python setup.py develop under _submodules/mmdetection.

(Assuming all datasets are stored in ~/data) Distributed training:

sh tools/dist_train_qg_rcnn.sh

Non-distributed training:

python tools/train_qg_rcnn.py --config configs/qg_rcnn_r50_fpn.py --load_from checkpoints/qg_rcnn_r50_fpn_2x_20181010-443129e1.pth --gpus 1

Before train, you'll need to download the initial weights transferred from FasterRCNN (provided by mmdetection, pretrained on COCO) to start.

Change the arguments in dist_train_qg_rcnn.sh or append them to python tools/train_qg_rcnn.py for your need. See train_qg_rcnn.py for details.

(Assuming all datasets are stored in ~/data).

python tools/test_global_track.py

Change the parameters, such as cfg_file, ckp_file and evaluators in test_global_track.py for your need.

• Initial weights transferred from FasterRCNN:

  • Google Drive: https://drive.google.com/open?id=1JkOqbSQJvGiGb9ubFIu5M84jeTInFJTX
  • Baidu Yun: [link] https://pan.baidu.com/s/1roviZxaV-A4QpS7FFSYT0g [password] 38qo

• Pretrained GlobalTrack:

  • Google Drive: https://drive.google.com/open?id=1ouTJDS5NACLukd810uiKCvmLxJfga48x
  • Baidu Yun: [link] https://pan.baidu.com/s/19Z7vWceXeF1EEpsOW_V-dQ [password] 47p4

By defaults, all pretrained weights are saved at checkpoints.

Please report issues in this repo if you have any problems.