# Non-local Neural Networks

**Authors:** Xiaolong Wang, Ross Girshick, Abhinav Gupta, Kaiming He  
**arXiv:** [1711.07971](https://arxiv.org/abs/1711.07971)  
**Date:** 2017-11-21  
**Summary page:** https://www.summarizepaper.com/en/paper/1711.07971v1/

## Key points

- Authors Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He introduce non-local operations for capturing long-range dependencies in computer vision tasks
- Non-local operation computes response at a specific position by considering weighted sum of features from all positions
- Can be seamlessly integrated into various computer vision architectures
- Non-local models demonstrate effectiveness across different tasks:
- Competitive performance in video classification without additional enhancements on datasets like Kinetics and Charades
- Significant improvements in static image recognition tasks such as object detection/segmentation and pose estimation on COCO suite of tasks
- Introduction of non-local operations enhances neural networks' ability to capture long-range dependencies in visual data
- Authors plan to release code for non-local models for further research and application in computer vision tasks

## Summary

In their paper titled "Non-local Neural Networks," authors Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He introduce non-local operations as a novel approach for capturing long-range dependencies in computer vision tasks. Drawing inspiration from the classical non-local means method, the non-local operation computes the response at a specific position by considering a weighted sum of features from all positions. This innovative building block can be seamlessly integrated into various computer vision architectures. The authors demonstrate the effectiveness of non-local models across different tasks. In video classification, even without additional enhancements, the non-local models exhibit competitive performance compared to current state-of-the-art methods on datasets such as Kinetics and Charades. Moreover, in static image recognition tasks like object detection/segmentation and pose estimation on the COCO suite of tasks, the non-local models show significant improvements. Overall, the introduction of non-local operations presents a promising avenue for enhancing the ability of neural networks to capture long-range dependencies in visual data. The authors plan to release code for their non-local models, making it accessible for further research and application in computer vision tasks.