litepose
The project introduces LitePose, an advanced architecture for multi-person pose estimation tailored for edge devices. By refining HRNet-based models, it adopts a streamlined single-branch design enhancing both efficiency and performance. Methods such as Fusion Deconv Head and Large Kernel Convs are utilized to minimize redundancy and boost model capacity. LitePose achieves higher mAP with a 25% computation increase compared to earlier models and significantly lowers latency on mobile platforms, promoting real-time estimation without compromising performance.
