/AdaMOT

(IEEE Transactions on Image Processing, 2022) "A closer look at the joint training of object detection and Re-identification in multi-object tracking"

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

AdaMOT

Codes for the paper "A closer look at the joint training of object detection and Re-identification in multi-object tracking" (IEEE Transactions on Image Processing, 2022)

Alt text

Our main contributions, Identity-aware Label Assignment and Discriminative Focal Loss are implemented in ada_matcher.py and ada_loss.py.

The entire project is largely based on FairMOT (FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking), most of the upgraded ccomponments in AdaMOT are named as "ada_xxx.py".

Our results:
Alt text

Usage:
    (To train) python ada_train.py --exp_id adamot_17 --load_model detector_pretrained_weights.pth --data_cfg src/lib/cfg/mot17.json --batch_size 24 --lr_step 20 --pos 7 --id_aware
    (To test) python ada_track.py --load_model trained_weights.pth --conf_thres 0.4 --test_mot17 True
    Other usage is identitcal to FairMOT
Coming soon:

The pre-trained weights.

The codes and of pre-trained weights of our improved CenterNet detector inside AdaMOT.

Alt text