Reproduce “Group channel pruning and spatial attention distilling for object detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yun Chu, Pu Li, Yong Bai, Zhuhua Hu, Yongqing Chen, Jiafeng Lu
A three-stage compression method tailored for object detection networks enables efficient deployment on resource-constrained devices. It achieves up to 64.7% parameter reduction and 34.9% FLOPs reduction for YOLOv4 with only a 0.5% [email protected] drop, leading to up to 1.5x faster inference on edge hardware without specialized support.
Key insights
- Redundancy levels vary significantly across different functional components (groups) of an object detection network, necessitating differentiated pruning thresholds for optimal compression.
- Dynamically adjusting the sparse regularization rate during training offers a more efficient way to achieve high network sparsity while maintaining accuracy compared to using a constant sparse rate.
- Distilling spatial attention maps from multi-scale feature layers effectively recovers the accuracy of pruned object detection models, proving robust and generalizable across various pruning methods and backbone architectures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.