Enhanced Head: Exploring Strong Detection Heads with Vision Transformer

This repository presents the official PyTorch implementation of Enhanced Head.

In this paper, we propose the Enhanced Head, which simultaneously incorporates Contextual Information Utilization Capability, Deformation Adaptation Capability, and Alignment Capability into various detectors. By replacing the original Decoupled Head with the proposed Enhanced Head, we achieve a reduction in computational complexity while delivering substantial performance improvements across multiple detectors.

Here is the overall architecture of the proposed Enhanced Head.

Installation

Please see get_started.md for the basic usage of MMDetection.

Training

bash tools/dist_train.sh configs/EnhancedHead/atss_enhanced.py 4

Testing

python tools/test.py configs/EnhancedHead/atss_enhanced.py work_dirs/atss_enhanced/latest.pth --eval bbox

Weight

Model	AP	Link1	Link2
atss_r50_enhanced	42.3	BaiduNetDisk	GoogleDrive
retina_r50_enhanced	39.8	BaiduNetDisk	GoogleDrive
fcos_r50_enhanced	40.8	BaiduNetDisk	GoogleDrive

Acknowledgement

This repository is built upon the MMDetection library.