# Light-Head R-CNN: In Defense of Two-Stage Object Detector

**Authors:** Zeming Li, Chao Peng, Gang Yu, Xiangyu Zhang, Yangdong Deng, Jian Sun  
**arXiv:** [1711.07264](https://arxiv.org/abs/1711.07264)  
**Date:** 2017-11-20  
**Summary page:** https://www.summarizepaper.com/en/paper/1711.07264v2/

## Key points

- Authors investigate limitations of two-stage methods compared to single-stage detectors in terms of speed
- Faster R-CNN and R-FCN involve intensive computations after or before Region of Interest (RoI) warping
- Heavy-head designs in architecture contribute to slow speed of these networks
- Authors propose a new two-stage detector called Light-Head R-CNN with a lightweight head design
- Light Head R CNN outperforms state-of-the-art object detectors on COCO dataset while maintaining time efficiency
- Achieves impressive results by replacing backbone with smaller network such as Xception
- Achieves 30.7 mmAP at 102 FPS on COCO, surpassing YOLO and SSD in terms of both speed and accuracy
- Authors plan to make their code publicly available for further exploration and implementation

## Summary

In this paper titled "Light-Head R-CNN: In Defense of Two-Stage Object Detector," authors Zeming Li, Chao Peng, Gang Yu, Xiangyu Zhang, Yangdong Deng, and Jian Sun investigate the limitations of typical two-stage methods compared to single-stage detectors like YOLO and SSD in terms of speed. They identify that Faster R-CNN and R-FCN involve intensive computations after or before Region of Interest (RoI) warping. Faster R-CNN utilizes two fully connected layers for RoI recognition, while R-FCN produces large score maps. These heavy-head designs in the architecture contribute to the slow speed of these networks. Even reducing the base model significantly does not lead to a proportional decrease in computation cost. To address these shortcomings in current two-stage approaches, the authors propose a new two-stage detector called Light-Head R-CNN. The key idea behind their design is to make the head of the network as lightweight as possible. They achieve this by using a thin feature map and a cost effective R-CNN subnet consisting of pooling and a single fully connected layer. Their ResNet 101 based Light Head R CNN outperforms state of the art object detectors on COCO dataset while maintaining time efficiency. Importantly, by simply replacing the backbone with a smaller network such as Xception, Light Head R CNN achieves impressive results. It achieves 30.7 mmAP at 102 FPS on COCO significantly surpassing single stage fast detectors like YOLO and SSD in terms of both speed and accuracy. The authors plan to make their code publicly available for further exploration and implementation which will enable researchers to further improve upon their work. Overall their research provides insights into improving the speed and efficiency of two stage object detection methods while achieving competitive performance on benchmark datasets like COCO.