yihongxu/daumot

Official Implementation for DAUMOT: Domain Adaptation for Unsupervised Multiple Object Tracking, An unsupervised MOT training framework with domain adaptation.

This is a placeholder for DAUMOT code release. The code will be released in the following days.

Official Implementation for DAUMOT: Domain Adaptation for Unsupervised Multiple Object Tracking

** DAUMOT: Domain Adaptation for Unsupervised Multiple Object Tracking **

The idea was originated from Guillaume Delorme and he implemented the code for FairMOT-DAUMOT, Yihong finished the TransCenter-DAUMOT part.

Guillaume Delorme*, Yihong Xu*, Luis G. Camara, Elisa Ricci, Radu Horaud, Xavier Alameda-Pineda

Sample qualitative results from different sequences of MOT20 comparing direct transfer (i.e. no adaptation) to DAUMOT. (a) shows the persons missed by direct transfer whereas (b) showcases over detections and (c,d) show person identity changes. The images illustrate the existence of large domain shifts between source (MOT17) and target (MOT20) when the model is not trained with DAUMOT. Green arrows point to the errors.