I am currently experimenting with Torchreid (using osnet_ibn_x1_0) for a small person re-identification project.
I would like to clarify how robust the model is to appearance changes. Specifically, I am interested in understanding:
- How much does OSNet rely on clothing appearance for matching?
- How robust is it to changes in clothing or accessories?
- Does the model capture more identity-related cues (e.g., body shape, proportions, gait-related patterns), or is it primarily appearance-driven?
- Would fine-tuning on a custom dataset with clothing variation significantly improve robustness to such changes?
From my initial experiments, matching performance appears to decrease when clothing changes, so I would like to better understand whether this is expected behavior for standard ReID models.
Thank you.