Skip to content

Latest commit

 

History

History

TraDeS

Track to Detect and Segment: An Online Multi-Object Tracker (CVPR 2021)

Citation

The code is built with TraDeS. Thanks for their great work.

Installation

Please refer to INSTALL.md for installation instructions.

Data preparation

The data is in the following structure:

DIVOTrack
    └——————datasets
    |        └——————DIVO
    |           |——————images
    |           |        └——————train
    |           |        └——————test
    |           └——————labels_with_ids
    |                    └——————train
    |                    └——————test  
    └——————${ROOT}

Train

Download the pre-trained model and put it into ./models/.

cd ${TraDeS_ROOT}
cd src
python main.py tracking --exp_id trades_divo --dataset divo --pre_hm --ltrb_amodal --same_aug --hm_disturb 0.05 --lost_disturb 0.4 --fp_disturb 0.1 --gpus 0,1 --save_point 10,20,25,30 --load_model ../models/crowdhuman.pth --clip_len 3 --max_frame_dist 10  --batch_size 12 --trades

or

sh ./experiments/divo_train.sh

Inference

Download our final model. Make sure the "exp_id" is the same as the training one, the model will be directly loaded from the corresponding dir "exp/tracking/"exp_id"/model_last.pth".

cd ${TraDeS_ROOT}
cd src
python test.py tracking --exp_id trades_divo --dataset divo --pre_hm --ltrb_amodal --pre_thresh 0.5 --inference --clip_len 3 --track_thresh 0.4 --gpus 0 --trades --resume

or

sh ./experiments/divo_test.sh

The result will be saved to "exp/tracking/"exp_id"/result_divo

Evaluation

  1. Change the directory name from "result_divo" to "trades"
  2. Make sure "centertrack" has the middle directory "data". (i.e. trades/data/circleRegion_Drone.txt instead of trades/circleRegion_Drone.txt)
  3. Copy your result_divo to DIVOTrack/TrackEval/data/trackers/mot_challenge/divo
  4. Go to DIVOTrack/TrackEval
  5. See the instrcution on TrackEval

Acknowledgment

Many thanks to CenterTrack authors for their great framework!