songsanling/TNT

TNT

Environment setting is follows.
Operating system: windows, need to change (from scipy.io import loadmat in TNT/train_cnn_trajectory_2d.py) to (h5py) in Linux.
Python: 3.5.4
tensorflow: 1.4.0
cuda: 8.0
cudnn: 5.1.10
opencv: 3.2.0
Other packages: numpy, pickle, sklearn, scipy, matplotlib, PIL.

2D Tracking Training

1. Prepare the Data.
   Ground truth tracking file: follow the format of MOT (https://motchallenge.net/).
   The frame index and object index are from 1 (not 0) for both tracking ground truth and video frames.
   
2. Convert MOT fromat to UA-Detrac format.
   TNT/General/MOT_to_UA_Detrac.m
   
3. Crop the ground truth detection into individual bounding box images.
   TNT/General/crop_UA_Detrac.m
   
4. Create validation pairs for FaceNet.
   TNT/General/create_pair.m
   
5. Train the triplet appearance model based on FaceNet using the cropped data.
   See https://github.com/davidsandberg/facenet.
   All the useful scource code are in TNT/src/.
   
6. Train 2D tracking.
   Set directory paths in TNT/train_cnn_trajectory_2d.py before the definition of all the functions.
   Change the sample probability (sample_prob) according to your data density. The number of element in sample_prob is the number of your input Mat files.
   Set the learning rate (lr) to 1e-3 at the beginning. Every 2000 steps, decrease lr by 10 times until it reaches 1e-5.
   The output model will be stored in save_dir.
   
7. Run python TNT/train_cnn_trajectory_2d.py.
   

2D Tracking Testing

1. Prepare the detection data.
   follow the format of MOT (https://motchallenge.net/).
   The frame index and object index are from 1 (not 0) for both tracking ground truth and video frames.
   
2. Set your data and model paths correctly on the top of TNT/tracklet_utils_3c.py.
   
3. Set the file_len to be the string length of your input frame name before the extension.
   
4. Adjust the tracking parameters in track_struct['track_params'] of TNT/tracklet_utils_3c.py in the function TC_tracker().
   
5. Run python TNT/TC_tracker.py.
   

Citation

Use this bibtex to cite this repository:

@article{wang2018exploit,
  title={Exploit the Connectivity: Multi-Object Tracking with TrackletNet},
  author={Wang, Gaoang and Wang, Yizhou and Zhang, Haotian and Gu, Renshu and Hwang, Jenq-Neng},
  journal={arXiv preprint arXiv:1811.07258},
  year={2018}
}