# Make Skeleton-based Action Recognition Model Smaller, Faster and Better

**Authors:** Fan Yang, Sakriani Sakti, Yang Wu, Satoshi Nakamura  
**arXiv:** [1907.09658](https://arxiv.org/abs/1907.09658)  
**Date:** 2019-07-23  
**Summary page:** https://www.summarizepaper.com/en/paper/1907.09658v1/

## Key points

- Skeleton-based action recognition has seen advancements but faces challenges like large model sizes and slow execution speeds.
- Researchers led by Fan Yang, Sakriani Sakti, Yang Wu, and Satoshi Nakamura proposed the Double-feature Double-motion Network (DD-Net) to address these issues.
- DD-Net utilizes a lightweight network structure with approximately 0.15 million parameters for remarkable speed performance.
- DD-Net achieves up to 3,500 frames per second on a single GPU or 2,000 FPS on a CPU.
- DD-Net leverages robust features for accurate action recognition and has demonstrated state-of-the-art performance on datasets like SHREC and JHMDB.
- The researchers plan to release the code associated with their work alongside the publication of their paper.

## Summary

In recent years, skeleton-based action recognition has seen significant advancements. However, many existing methods face challenges such as large model sizes and slow execution speeds. To address these issues, a team of researchers led by Fan Yang, Sakriani Sakti, Yang Wu, and Satoshi Nakamura have proposed a novel approach called the Double-feature Double-motion Network (DD-Net). By analyzing the properties of skeleton sequences, DD-Net utilizes a lightweight network structure with approximately 0.15 million parameters. This design enables DD-Net to achieve remarkable speed performance, reaching up to 3,500 frames per second (FPS) on a single GPU or 2,000 FPS on a CPU. One of the key strengths of DD-Net lies in its ability to leverage robust features for accurate action recognition. Through experimental evaluations on datasets such as SHREC (focused on hand actions) and JHMDB (centered around body actions), DD-Net has demonstrated state-of-the-art performance levels. The researchers plan to release the code associated with their work alongside the publication of their paper. Overall, the innovative approach presented in this study not only addresses the limitations of existing skeleton-based action recognition methods but also sets new benchmarks in terms of model efficiency and speed without compromising on accuracy.