Awesome

Deep-TAMA 2021

The paper was accepted in Elsevier Information Sciences (IF 6.795)

• arXiv paper link : https://arxiv.org/abs/1907.00831

Enviroment

OS : Windows10 64bit (Verified to works fine on Ubuntu 18.04)
CPU : Intel i5-8500 3.00GHz
GPU : Geforce GTX Titan X (Works on GPU with smaller memory size >= 5GB)
RAM : 32 GB

Requirements

python 3.6
tensorflow-gpu 2.1.0 (strict!)
numpy 1.17.3
opencv 3.4.2
matplotlib 3.1.1
scikit-learn 0.22.1
motmetrics

Sample tracking dataset structure

- Set the dataset folder as following structure
  ex) MOT
      |__ TUD-Stadtmitte
      |         |__ det
      |         |__ gt
      |         |__ img1
      |__ MOT16-02
      |__ Custom_Seuqnce
      .
      .
- We recommend to copy-and-paste all MOTChallenge sequences in MOT folder
  

Tracking settings

1. Download the pre-trained models and locate in in model directory.

2. Set the variable seq_path in config.py to your own dataset path.

   • Dataset should be 'MOT/{sequence_folder-1, ..., sequence_folder-N}'.
   • Each of sequence_folder should follow the MOTChallenge style (e.g., 'sequence_folder-1/{det, gt, img1}').
   • The simplest way is just copy and paste all MOTChallenge datasets (2DMOT2015, MOT16, MOT16, MOT20, etc) in MOT folder.
   • The compatible datasets are available on MOTChallenge.

3. Set the variable seqlist_name in racking_demo.py to the proper name.

   • Exemplar sequence groups are provided.
   • Add your own tracking sequence group in sequence_groups.

4. Perform tracking using tracking_demo.py.

   • Tracking thresholds can be controlled by modifying config.py.
   • List of significant commandline arguments in tracking_demo.py.
     • set_fps : manipulates an FPS and drop frames of videos
     • semi_on : improves a tracking performance using interpolation and restoration
     • init_mode : mht (faster, geometry-based), delayed (geometry + appearance)
     • gating_mode : iou (intersection-over-union), maha (pose + shape)
     • only_motion : Do not use appearance matching

Training settings

1. Set the data as same as 'Tracking settings' above.

2. Modify the sequence_groups/trainval_group.json to your own dataset

   • Training and validation dataset should have gt folder.

3. Perform training using training_demo.py.

   • JI-Net training should be done first.
   • Using JI-Net model as a feature extractor, LSTM can be trained.

Evaluation

• Evaluated performance on sequence group validation.txt.
  • The highlighted performance can be achieved by default demo setting.

• For evaluation, set the command line argument --evaluate in tracking_demo.py
• Currently, our code doesn't support HOTA. For HOTA, check TrackEval.
• The code produces tracking results in both txt and image format.

Pre-trained models

• Validation accuracy and loss comparison
  • Both of JI-Net and LSTM used cross-entropy loss for training
  <p float="left"> <img src="/images/acc_comparison.png" height="200"> <img src="/images/loss_comparison.png" height="200"> </p>
• We provide pre-trained models for JI-Net and LSTM
  • Locate the downloaded models in model directory
  • Download links
    • JI-Net : https://drive.google.com/file/d/1Lreh8FxDYYx3ymgm9B4TuD_lz0y7Vi7C/view?usp=sharing
    • LSTM : https://drive.google.com/file/d/1KbUHnfVSRkuV6SMFCdvs7U5PuRoR2Dxi/view?usp=sharing

Qualitative results

Reference

@inproceedings{ycyoon2018,
  title={Online Multi-Object Tracking with Historical Appearance Matching and Scene Adaptive Detection Filtering},
  author={Young-Chul Yoon and Abhijeet Boragule and Young-min Song and Kwangjin Yoon and Moongu Jeon},
  year={2018},
  booktitle={IEEE AVSS}
}
@inproceedings{ycyoon2021,
  title={Online Multiple Pedestrians Tracking using Deep Temporal Appearance Matching Association},
  author={Young-Chul Yoon Du Yong Kim and Young-min Song and Kwangjin Yoon and Moongu Jeon},
  year={2021},
  booktitle={Information Sciences}
}

• This tracker has been awarded a 3rd Prize on 4th BMTT MOTChallenge Workshop held in CVPR 2019
• If you have any question, feel free to contact by yyc9268@gmail.com