A Multi-scale Feature Embedding Extension Network for RGB-D Cross-modal Person Re-identification

The RGB-D cross-modal person re-identification task is a branch of the cross-modal person re-identification task, aiming to match pedestrian images between RGB and depth modalities. Its main challenge lies in the significant modality difference between RGB and depth images. Due to the limited research attention on this task, there is a lack of large-scale, high-quality training data, making existing methods struggle to capture effective correlational information between RGB and depth modalities. This paper proposes a novel feature embedding expansion network for RGB-D cross-modal person re-identification. The network introduces perturbations to the embedding of intermediate-layer features to generate multiple diversified embedding features, effectively alleviating the problem of limited training data for RGB-D tasks. Simultaneously, through the incorporation of deformable convolution operations and depth-wise separable convolution operations during the embedding feature generation, the model is encouraged to learn richer scale and channel information from the limited training data. Meanwhile, by controlling the distance between expanded feature embeddings and original feature embeddings for different modalities, we mitigate the significant modality differences. Experimental results on the public benchmark datasets BIWI and RobotPKU demonstrate that our approach outperforms state-of-the-art methods.