Pose-Transformation and Radial Distance Clustering for Unsupervised Person Re-identification

TL;DR

This work proposes an unsupervised approach to the person re-ID setup with zero knowledge of true labels, and introduces a novel radial distance loss, that attends to the fundamental aspects of feature learning - compact clusters with low intra-cluster and high inter-cluster variation.

Abstract

Person re-identification (re-ID) aims to tackle the problem of matching identities across non-overlapping cameras. Supervised approaches require identity information that may be difficult to obtain and are inherently biased towards the dataset they are trained on, making them unscalable across domains. To overcome these challenges, we propose an unsupervised approach to the person re-ID setup. Having zero knowledge of true labels, our proposed method enhances the discriminating ability of the learned features via a novel two-stage training strategy. The first stage involves training a deep network on an expertly designed pose-transformed dataset obtained by generating multiple perturbations for each original image in the pose space. Next, the network learns to map similar features closer in the feature space using the proposed discriminative clustering algorithm. We introduce a novel radial distance loss, that attends to the fundamental aspects of feature learning - compact clusters with low intra-cluster and high inter-cluster variation. Extensive experiments on several large-scale re-ID datasets demonstrate the superiority of our method compared to state-of-the-art approaches.

Figures (3)

• Figure 1: Pipeline of the proposed person re-id framework. PT Dataset: Poses and parts corresponding to the original dataset are obtained. Parts are transformed spatially as per the sampled poses, $p^o_t$ to generate images, $x_i^p$. PT dataset $=$ original dataset $+$ generated samples is used to train the initialization stage. Initialization Stage: Trained using triplet and classification losses. Discriminative Clustering: Weights for $\mathcal{F}$ are taken from the Initialization Stage. $\mathcal{F}$ is trained only on the original dataset using triplet, classification losses and the proposed radial distance loss. $\gamma$ represents the minimum radial distance. The blue point violates $\gamma$ for Orange cluster, and thus, is pushed away.
• Figure 2: t-SNE visualization for A. 5 clusters with similar appearing but different identities, represented by images in the top row. B. 20 randomly chosen clusters
• Figure 3: Ablation on the Market-1501 by varying the margin $\gamma$, and sample scaling factor $K$.