Expanding and Refining Hybrid Compressors for Efficient Object Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Lightweight models trained via knowledge distillation (KD) are efficient for object re-identification (Re-ID).
• Architectural differences between student and teacher models hinder knowledge transfer, impacting accuracy.

Purpose of the Study:

• To propose a novel method that enhances the representational capacity of lightweight Re-ID models while pruning complexity.
• To improve the efficiency and accuracy of object re-identification models.

Main Methods:

• Introduced a refiner-expander-refiner (RER) structure with a multi-branch convolutional expander to increase student model capacity.
• Implemented 1x1 convolutional layers as refiners to prune expander input/output channels.
• Developed a common consensus gradient resetting (CCGR) method to manage accuracy and pruning gradients.
• Utilized re-parameterization to simplify the trained RER into a slim convolutional layer for faster inference.

Main Results:

• The proposed Expanding and Refining Hybrid Compressing (ERHC) method demonstrated superior inference speed and accuracy.
• On the VeRi-776 dataset, ERHC achieved significant reductions: 75.33% fewer model parameters (MP) and 74.29% fewer floating-point operations (FLOPs) when using ResNet101 as a teacher.
• Accuracy was maintained despite substantial model compression.

Conclusions:

• The ERHC method effectively addresses the knowledge transfer challenges in KD for Re-ID.
• ERHC offers a viable approach for creating highly efficient and accurate lightweight Re-ID models.
• The RER structure and CCGR method contribute to improved model compression and performance.