search

layumi / Person_reID_baseline_pytorch

:bouncing_ball_person: Pytorch ReID: A tiny, friendly, strong pytorch implement of person re-id / vehicle re-id baseline. Tutorial 👉https://github.com/layumi/Person_reID_baseline_pytorch/tree/master/tutorial
https://www.zdzheng.xyz
MIT License
4.13k stars 1.01k forks source link

Clarification needed #376

Closed robosina closed 1 year ago

robosina commented 1 year ago

I'm curious if person re-identification (ReID) can be used like face recognition tasks. I noticed that we get an output, for example, a vector of size 751 which matches the number of people in the dataset. But, can we use a technique like metric learning here? This would let us use an image of a person as a reference, and then for each new frame, we get an 'embedding vector' for each person. We could then compare these vectors to the reference using cosine similarity.

What I really want to know is, can this person re-identification method in the repository work with people it hasn't seen before if we provide a reference image? In other words, I'm looking for a more general person re-identification method that can handle new people, not just the ones in the original dataset.

layumi commented 1 year ago

Yes. But it requires a large and general dataset.

You may refer to our latest work https://github.com/Shuyu-XJTU/APTM It is pre-trained on a large dataset generated by Stable Diffusion.

robosina commented 1 year ago

Thanks