CCPA: Long-term Person Re-Identification via Contrastive Clothing and Pose Augmentation

TL;DR

This work proposes CCPA: Contrastive Clothing and Pose Augmentation framework for LRe-ID, and performs clothing and pose transfer across identities to generate images of more clothing variations and of different persons wearing similar clothing.

Abstract

Long-term Person Re-Identification (LRe-ID) aims at matching an individual across cameras after a long period of time, presenting variations in clothing, pose, and viewpoint. In this work, we propose CCPA: Contrastive Clothing and Pose Augmentation framework for LRe-ID. Beyond appearance, CCPA captures body shape information which is cloth-invariant using a Relation Graph Attention Network. Training a robust LRe-ID model requires a wide range of clothing variations and expensive cloth labeling, which is lacked in current LRe-ID datasets. To address this, we perform clothing and pose transfer across identities to generate images of more clothing variations and of different persons wearing similar clothing. The augmented batch of images serve as inputs to our proposed Fine-grained Contrastive Losses, which not only supervise the Re-ID model to learn discriminative person embeddings under long-term scenarios but also ensure in-distribution data generation. Results on LRe-ID datasets demonstrate the effectiveness of our CCPA framework.

Figures (3)

• Figure 1: (a) Long-term Re-ID aims to re-identifies the same person under different clothing, viewpoint, illumination, etc.; (b) t-SNE visualization of distribution in the latent space of person embeddings output by the model trained without the proposed losses. Viewpoint variations cause severe ambiguity in learning pose-based shape, leading to large intra-class and small inter-class gap, shown by the widespread of red circles. This forms the motivation of this paper.
• Figure 2: Overview of the proposed CCPA framework. From the original batch, for every pair of images, clothing and pose transfer is performed using an appearance encoder $E^A$ (a CNN), a shape encoder $E^S$ and a decoder $G$ to form the augmented batch. Contrastive sampling is performed on the augmented batch for inputs to the proposed fine-grained contrastive losses, which drive model training along with an identification loss. (Best viewed in color)
• Figure 3: Architecture of Shape Encoder, which comprises of a pose estimator, a refinement network and a R-GAT.