LSH9832/edgeyolo
an edge-real-time anchor-free object detector with decent performance
What it solves
EdgeYOLO is designed to provide high-performance object detection specifically for embedded and edge devices. It addresses the trade-off between detection accuracy and inference speed, enabling real-time processing on hardware with limited computational resources, such as the Nvidia Jetson AGX Xavier.
How it works
EdgeYOLO is an anchor-free object detector. It utilizes a range of model sizes (from Tiny to L) to fit different hardware constraints. The project includes a specialized RH loss function used in the final stages of training to improve detection performance for small and medium models. It also provides enhanced data augmentation methods to maintain effectiveness even when dataset labels are sparse.
Who it’s for
This project is for developers and researchers working on computer vision for edge computing, robotics, and embedded systems who need a fast, accurate, and easily deployable object detection model.
Highlights
- Edge-Optimized Performance: Achieves up to 34 FPS on Nvidia Jetson AGX Xavier with 50.6% AP on COCO2017.
- Broad Deployment Support: Includes tools and code for exporting models to ONNX, TensorRT (FP16 and INT8), RKNN, MNN, Horizon J5, and Huawei Ascend (OM).
- ** uma Unified Deployment Branch**: A dedicated deployment branch simplifies multi-platform deployment without requiring code changes per platform.
- Flexible Dataset Support: Compatible with COCO, YOLO, VOC, VisDrone, and DOTA dataset formats.
- Comprehensive Tooling: Provides Docker images for one-click model export and scripts for plotting loss, learning rate, and accuracy curves.