Learning Intra and Inter-Camera Invariance for Isolated Camera Supervised Person Re-identification

Menglin Wang,Xiaojin Gong

2023-11-02

Abstract:Supervised person re-identification assumes that a person has images captured under multiple cameras. However when cameras are placed in distance, a person rarely appears in more than one camera. This paper thus studies person re-ID under such isolated camera supervised (ISCS) setting. Instead of trying to generate fake cross-camera features like previous methods, we explore a novel perspective by making efficient use of the variation in training data. Under ISCS setting, a person only has limited images from a single camera, so the camera bias becomes a critical issue confounding ID discrimination. Cross-camera images are prone to being recognized as different IDs simply by camera style. To eliminate the confounding effect of camera bias, we propose to learn both intra- and inter-camera invariance under a unified framework. First, we construct style-consistent environments via clustering, and perform prototypical contrastive learning within each environment. Meanwhile, strongly augmented images are contrasted with original prototypes to enforce intra-camera augmentation invariance. For inter-camera invariance, we further design a much improved variant of multi-camera negative loss that optimizes the distance of multi-level negatives. The resulting model learns to be invariant to both subtle and severe style variation within and cross-camera. On multiple benchmarks, we conduct extensive experiments and validate the effectiveness and superiority of the proposed method. Code will be available at <a class="link-external link-https" href="https://github.com/Terminator8758/IICI" rel="external noopener nofollow">this https URL</a>.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to reduce the impact of camera - style deviation on cross - camera pedestrian recognition in the Person Re - Identification (Re - ID) task under the Isolated Camera Supervision (ISCS) setting. Specifically, when the distance between cameras is large, the same person rarely appears in multiple cameras simultaneously, which makes it difficult to apply traditional cross - camera pedestrian re - identification methods. In this case, each pedestrian has only a limited number of images from a single camera, so camera - style deviation becomes a key factor affecting the distinction of pedestrian identities. Cross - camera images may be misidentified as different pedestrian identities simply because of differences in camera style.

To eliminate the confounding effect of camera - style deviation, the paper proposes a new perspective, that is, learning intra - camera and cross - camera invariance within a unified framework. This is mainly achieved through the following methods:

1. **Construct a style - consistent environment**: Through clustering techniques, construct a style - consistent environment and perform prototype contrastive learning in each environment.
2. **Enhance intra - camera invariance**: Through strong data augmentation, contrast with the original prototype to strengthen the enhanced invariance within the camera.
3. **Improve multi - camera negative sample loss**: Design an improved multi - camera negative sample loss to optimize the distance of multi - level negative samples, thereby improving cross - camera invariance.

Through these methods, the model can learn features that are invariant to both subtle and severe style changes, thus verifying its effectiveness and superiority in multiple benchmark tests.