# Point Linking Network for Object Detection

**Authors:** Xinggang Wang, Kaibing Chen, Zilong Huang, Cong Yao, Wenyu Liu  
**arXiv:** [1706.03646](https://arxiv.org/abs/1706.03646)  
**Date:** 2017-06-12  
**Summary page:** https://www.summarizepaper.com/en/paper/1706.03646v1/

## Key points

- Object detection is a fundamental problem in computer vision
- Recent advancements in deep Convolutional Neural Networks (ConvNets) have improved object detectors
- Most existing deep ConvNet-based object detectors focus on regressing bounding box coordinates
- The proposed approach, Point Linking Network (PLN), represents bounding boxes using points and links
- PLN uses a fully convolutional network to regress corner/center points of the bounding box and their corresponding links
- By mapping these points and links back to multiple bounding boxes, PLN obtains an object detection result by fusing them
- PLN is robust to object occlusion and flexible to scale and aspect ratio variation due to its point-based representation and link connections within deep ConvNets
- PLN achieves state-of-the-art performance on benchmark datasets such as PASCAL VOC 2007, PASCAL VOC 2012, and COCO detection benchmarks using the Inception-v2 model
- The source code for PLN will be made publicly available for further research.

## Summary

Object detection is a fundamental problem in computer vision and recent advancements in deep Convolutional Neural Networks (ConvNets) have significantly improved the performance of object detectors. Most existing deep ConvNet-based object detectors focus on regressing the coordinates of bounding boxes. To address this limitation, we propose a novel approach called Point Linking Network (PLN) for object detection. Instead of considering the bounding box as a single entity, our method represents it using points and links. We use a fully convolutional network to regress the corner/center points of the bounding box and their corresponding links. By mapping these corner points and links back to multiple bounding boxes, we obtain an object detection result by fusing these multiple bounding boxes. PLN is robust to object occlusion and flexible to object scale variation and aspect ratio variation due to its point-based representation and link connections within deep ConvNets. We evaluate PLN with the Inception-v2 model on popular benchmark datasets such as PASCAL VOC 2007, PASCAL VOC 2012, and COCO detection benchmarks. Our results demonstrate that PLN achieves state-of-the-art performance without any additional enhancements or modifications. The source code for PLN will be made publicly available to facilitate further research in this area.