Identity Hides in Darkness: Learning Feature Discovery Transformer for Nighttime Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Daytime person re-identification (Re-ID) methods struggle in low-light conditions due to degraded image quality.
• Existing nighttime Re-ID approaches using synthetic data or enhancement-then-match strategies face limitations in realism and feature discriminability.

Purpose of the Study:

• To propose a novel nighttime person Re-ID method, the Feature Discovery Transformer (FDT), designed to capture pedestrian identity information in darkness.
• To address the challenges of noise and irrelevant features in low-light person Re-ID.

Main Methods:

• The FDT model incorporates a Frequency-wise Reconstruction Module (FRM) using Discrete Haar Wavelet Transform to separate and learn from high- and low-frequency components.
• A Normalized Contrastive Loss (NCL) is proposed to prevent over-smoothing and enhance identity details in high-frequency features.
• An Attribute Guide Module (AGM) integrates auxiliary information (camera ID, viewpoint) to reduce bias from irrelevant features and improve robustness.

Main Results:

• The FDT model demonstrated state-of-the-art performance on two realistic nighttime person Re-ID benchmarks: Night600 and RGBNT201rgb.
• The proposed modules effectively reduce noise disturbance and extract discriminative identity features from low-light images.

Conclusions:

• The Feature Discovery Transformer (FDT) offers a significant advancement in nighttime person Re-ID.
• The method's ability to discover and leverage hidden identity details in darkness provides a robust solution for challenging low-light scenarios.