Feature-Tuning Hierarchical Transformer via token communication and sample aggregation constraint for object...

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Transformer-based methods are successful in object re-identification.
• Current methods treat all patch tokens equally, limiting feature representation.

Purpose of the Study:

• To develop a feature-tuning mechanism for transformer backbones to improve object re-identification.
• To emphasize important patches and attenuate unimportant ones for more discriminative features.

Main Methods:

• Propose a plug-and-play Feature-tuning module via Token Communication (TCF) within transformer encoder blocks.
• Construct a Feature-tuning Hierarchical Transformer (FHTrans) with three hierarchies for feature extraction.
• Introduce a Sample Aggregation (SA) loss for enhanced intra-class aggregation.

Main Results:

• The proposed FHTrans method achieves state-of-the-art performance on object re-identification benchmarks.
• The feature-tuning mechanism effectively emphasizes discriminative features.
• Hierarchical structure and SA loss contribute to improved feature learning.

Conclusions:

• The FHTrans model offers a significant advancement in object re-identification.
• The feature-tuning mechanism and hierarchical design are crucial for performance.
• The proposed method provides a more effective approach to learning discriminative features.