# SlowFast Networks for Video Recognition

**Authors:** Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, Kaiming He  
**arXiv:** [1812.03982](https://arxiv.org/abs/1812.03982)  
**Date:** 2018-12-10  
**Summary page:** https://www.summarizepaper.com/en/paper/1812.03982v1/

## Key points

- SlowFast networks consist of a Slow pathway and a Fast pathway
- Slow pathway captures spatial semantics at a low frame rate
- Fast pathway captures motion at a high frame rate
- Fast pathway can be made lightweight while still learning useful temporal information
- Models achieve strong performance in action classification and detection tasks
- Accuracy of 79.0% on Kinetics dataset without pre-training, surpassing previous best results
- Achieve new state-of-the-art mean average precision (mAP) of 28.3 on AVA action detection
- Combining slow and fast pathways effectively captures spatial and temporal information for accurate video recognition
- Code will be made publicly available for other researchers to reproduce and build upon the work
- SlowFast networks offer promising approach with significant performance gains compared to existing methods

## Summary

The paper introduces SlowFast networks for video recognition, which consist of a Slow pathway and a Fast pathway. The Slow pathway operates at a low frame rate to capture spatial semantics, while the Fast pathway operates at a high frame rate to capture motion at fine temporal resolution. The Fast pathway can be made lightweight by reducing its channel capacity but still learns useful temporal information for video recognition. The authors demonstrate that their models achieve strong performance in both action classification and detection tasks. They report an accuracy of 79.0% on the Kinetics dataset without using any pre-training, surpassing previous best results in this area. Additionally, they achieve a new state-of-the-art mean average precision (mAP) of 28.3 on AVA action detection. The paper highlights the contributions of the SlowFast concept in achieving these improvements. By combining the strengths of both slow and fast pathways, the model effectively captures spatial and temporal information for accurate video recognition. This allows it to better understand complex motion patterns and recognize actions more accurately than existing methods. The authors also mention that they will make their code publicly available, allowing other researchers to reproduce and build upon their work. Overall, the SlowFast networks presented in this paper offer a promising approach for video recognition tasks with significant performance gains compared to existing methods. With its ability to capture both spatial semantics and motion information effectively, it is well suited for applications such as action classification and detection where accuracy is paramount.