UT Arlington's GCP Beats Person Re-ID Records by Learning Smarter Gallery Prototypes
A new attention-based method generates multiple class prototypes per person, pushing person re-identification accuracy past 97% on Market-1501.
Takeaways
- New method GCP generalizes person re-identification by using multiple learned prototypes per identity instead of one centroid.
- A transformer decoder generates prototypes autoregressively, conditioned on all gallery images and camera IDs of a person.
- Achieves 97.3/97.1 R-1/mAP on Market-1501 and 93.1/92.2 on CUHK03-NP, beating prior state of the art.
- Training-free baseline alpha-FPS also beats prior SOTA mAP, showing the prototype framing matters as much as the model.
- Number of prototypes N is tunable: higher N boosts rank-1 accuracy but lowers mean average precision.
- Code available on GitHub; paper on arXiv.
Person re-identification systems have gotten very good at extracting features from single images, but they still struggle with a subtler question: when you have dozens of pictures of the same person from different cameras, what is the best way to represent that person for retrieval? A new paper from The University of Texas at Arlington argues that the answer sits between one centroid and every instance, and it uses a transformer decoder to figure out exactly where.
The work, called Person Re-Identification via Generalized Class Prototypes (GCP), reframes retrieval as a prototype-selection problem and shows measurable gains on three standard benchmarks. The source code is publicly available.
The overfitting-underfitting trap
Person re-identification (Re-ID) is a specific image retrieval problem where, given a query image of a person, the goal is to identify it from a set of gallery images captured by non-overlapping cameras. It is used in surveillance, sports analytics, and autonomous vehicles, and it is hard because lighting, poses, occlusions, and camera viewpoints vary wildly across the gallery.
Most modern pipelines convert each image to a feature vector and then compare distances. Two dominant approaches structure the gallery side of that comparison, and the authors argue both are flawed:
- Instance-based retrieval: keep every gallery vector and compare to the nearest one. The hypothesis space has a Vapnik-Chervonenkis dimension of O(n) where n is the gallery size, and in Re-ID tasks where n is much larger than 100, this makes the decision boundary susceptible to overfitting, reducing mean precision.