Cross-modal distillation for RGB-depth person re-identification

Person re-identification is a key challenge for surveillance across multiple sensors. Prompted by the advent of powerful deep learning models visual recognition, and inexpensive RGB-D cameras sensor-rich mobile robotic platforms, e.g. self-driving vehicles, we investigate relatively unexplored problem cross-modal persons between RGB (color) depth images. The considerable divergence in data distributions different sensor modalities introduces additional challenges to typical difficulties like distinct viewpoints, occlusions, pose illumination variation. While some work has investigated infrared, take inspiration from successes transfer object detection tasks. Our main contribution novel method distillation robust person re-identification, which learns shared feature representation space person’s appearance both In addition, propose attention mechanism where gating signal one modality can dynamically activate most discriminant CNN filters other modality. proposed compared conventional approaches cross-domain Results obtained on public BIWI RobotPKU datasets indicate that significantly outperform state-of-the-art up 16.1% mean Average Precision (mAP), demonstrating benefit paradigm. experimental results also using allows improve recognition accuracy considerably with respect relevant approaches. 1 Code: https://github.com/frhf/cross-modal-distillation-reidentification . • Cross-modal training procedure embedding Exploiting intrinsic relation RGB. Ideal needs place Novel gated improves our approach.