Low-rank Fusion Network for Multi-modality Person Re-identification
Ziling He, Hanqin Shi, Youqing Wu, Zhengzheng Tu · 2023
Ideally, multi-modality person re-identification (Re-ID) models can make use of the complementary information contained in different modality images to obtain richer target features under ideal conditions, thus making up for the shortcomings of traditional visible person re-identification in complex environments. However, the advantages of different modalities may not be fully exploited due to incomplete feature learning. Therefore, we propose a low-rank fusion network for multi-modality person Re-ID, which can extract multi-scale features of each modality under the guidance of efficient channel attention (ECA) to capture different details of the target image, and then apply the low-rank feature fusion module to achieve feature complementation between multiple modalities and learn recognition features with strong representational ability. Extensive experiments on multi-modality person Re-ID dataset RGBNT201 comparing to other advanced methods verify the effectiveness of our method in complex environments.