search

Zhongdao / Towards-Realtime-MOT

Joint Detection and Embedding for fast multi-object tracking
MIT License
2.38k stars 539 forks source link

MOT20 and IDS #118

Open samyhocine opened 4 years ago

samyhocine commented 4 years ago

Hi,

I'm a big fan of your work "Towards-Realtime-MOT" / "JDE" and I would like to know if you are planning on doing the MOT20 challenge and if you are working on reducing the IDS ?

Best regards,

Samy H.

Zhongdao commented 4 years ago

Thanks a lot for your interest in our work! I am afraid I cannot participate in the MOT20 challenge. My current work is about a better way to measure similarities between objects, it is related to JDE and may have potential on reducing the IDS, but it does not focus on the MOT task alone. I do have some plans to improve the JDE tracker. If our paper could be accepted, I guess I would spend more time on this project ...

samyhocine commented 4 years ago

Thanks for your feedback ;) Maybe you will be all set for MOT21 🐰

What do you mean by:

if our paper could be accepted

By the ANU ? I thought that your paper "Towards Real-Time Multi-Object Tracking" has already been accepted.

Have a nice day ;)

Zhongdao commented 4 years ago

This paper gets positive feedbacks from reviewers but is rejected by ACs in CVPR 2020, so that I submit it to ECCV 2020.

samyhocine commented 4 years ago

Oh :/ Sorry to hear that.

Do you know why they rejected it ? :/

Zhongdao commented 4 years ago

"The technical novelty is limited" and "The IDF score is not competitive". That's true, the idea is quite simple. If an idea is too straightforward, it is less likely to be accepted by conferences like CVPR, unless you get very impressive results. Maybe we need to pay more effort. However, we believe the joint solution of detection and MOT is one of the most promising directions for MOT. If this work could bring peers' attention to the joint solution, it's also a very good thing.