Answers are grounded in this paper and cite their sources.
Infrared (IR) person re-identification (Re-ID) supports surveillance in nighttime and low-light conditions, but low resolution, noise, and limited texture information hinder recognition. We propose a bio-inspired lightweight framework based on enhanced quantization-aware training-based knowledge distillation (EQT-KD). Selective feature suppression and adaptive weighting of head, upper-body, and lower-body regions provide a functional, design-level analogue of selective visual attention. Enhanced LightMBN serves as the teacher, while Extreme LightMBN is the compact student. EQT-KD adjusts branch-wise distillation strength according to the difference between student outputs before and after fake quantization, supporting stable knowledge transfer under quantization. Experiments on SYSU-MM01, DBPerson-Recog-DB1, and the near-infrared (NIR) subset of RGBNT201 yielded mean average precision values of 0.7646, 0.9883, and 0.7473 and Rank-1 accuracies of 0.8037, 0.9903, and 0.7547, respectively, under the IR-only protocol used in this study, in which each unseen test identity has one query and two gallery images. The student uses 2.35 million parameters and 0.66 GFLOPs. Inference-time measurements on desktop, NVIDIA Jetson Orin Nano, and mobile platforms indicate the potential for real-time processing. These results support selective feature processing and quantization-aware knowledge transfer as a practical approach to balancing recognition performance and computational efficiency in resource-constrained IR surveillance.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: