Reid model

[sky-creater]

Hello, thank you very much for your great work. I would like to ask if there are any experiments using FastReID as the ReID model here? Since many recent TBD methods adopt FastReID, would the absence of it in comparisons potentially lead to unfair experimental results?

[VlSomers]

Hi @sky-creater, no experiment rely on FastReID here, we only use some forks of Torchreid. We have however made our implementation quite modular and integrating FastReID should be a low hanging fruit: you can easily implement a wrapper that would serve as an interface to FastReID. About the fairness of the comparisons with other tracking methods utilising other ReId models, we have an extensive discussion on it in the supplementary materials of the paper, I invite you to read this section. Moreover, FastReID and Torchreid are two frameworks to train and serve ReID models, but both implement similar models with similar performance. A fair comparison is not about using the same ReID framework as other methods, but rather using the same ReID model, or at least a model with similar performance/size.

[bstandaert]

Hi @sky-creater,

You are right: ensuring fair comparisons across methods is a complex challenge, especially when it comes to ReID models. We did not use FastReID in our experiments for several reasons, which have been the subject of many internal discussions.

To be fully transparent, we have not yet identified a definitive approach that would make comparisons completely fair in this regard. This concern is something we have addressed in detail in our paper, available on arXiv. We specifically discuss that in Section D, under the paragraph titled “Re-ID Models”.

Feel free to take a look, and don’t hesitate to share your thoughts :)