Disentangled Representations for Short-Term and Long-Term Person Re-Identification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Person re-identification (reID) is challenging due to variations in appearance (pose, viewpoint) and shared attributes among different individuals.
• Existing reID methods often require specific supervisory signals (e.g., pose annotations) to learn discriminative features, limiting their generalizability.
• The need for robust person representations that are invariant to identity-unrelated factors is critical for effective reID.

Purpose of the Study:

• To develop a novel method for disentangling identity-related and identity-unrelated features in person images for improved reID.
• To propose a generative adversarial network, Identity Shuffle GAN (IS-GAN), that learns these disentangled representations using only identity labels.
• To validate the effectiveness of IS-GAN on standard reID benchmarks and demonstrate its advantage in long-term reID tasks.

Main Methods:

• Proposed Identity Shuffle GAN (IS-GAN), a generative adversarial network designed for feature disentanglement.
• Employed an identity-shuffling technique utilizing only person identification labels to separate identity-related and -unrelated features.
• Incorporated constraints on the distribution of identity-unrelated features and encouraged feature decorrelation to facilitate disentanglement.

Main Results:

• IS-GAN achieved state-of-the-art performance on widely-used reID benchmarks: Market-1501, CUHK03, and DukeMTMC-reID.
• Demonstrated superior performance on a long-term reID task, setting a new state-of-the-art on the Celeb-reID dataset.
• Validated the effectiveness of disentangling person representations for robust reID.

Conclusions:

• The proposed IS-GAN effectively disentangles identity-related and -unrelated features without auxiliary supervision, significantly advancing person reID.
• Feature disentanglement is a promising direction for improving the robustness and accuracy of person re-identification systems.
• The method offers a generalizable approach for learning discriminative person representations applicable to various reID scenarios.