Person Re-Identification Based on Multi-Spectrum Information Aggregation Transformers

Convolutional Neural Networks (CNNs) have achieved remarkable success in various image-related tasks. However, CNNs' chunking approach of sliding windows over pixels hinders its ability to learn sequential semantic information. To address this issue, researchers have proposed models that integrate Transformers to aggregate features across different resolutions, taking advantage of the Transformer's strengths. Hierarchical Aggregation Transformer (HAT) is one such model that integrates low-level and high-level details, showing a great effect on person re-identification. Based on HAT, we propose the Multi-Spectrum Information Aggregation Transformers (MIAT), which enhances the low-level structural information captured by CNNs through a frequency channel attention module that extracts multi-spectral features. By leveraging discrete cosine transform (DCT), more detailed features can be learned to improve person re-identification, enabling better discernment of distinguishing features between different individuals. Through extensive experimentation, we validate our model's performance on two widely employed person re-identification datasets and demonstrate its superiority over existing advanced methods.