YOLOX

Why we included this project

Most object detection work in production starts with a solid pretrained baseline, and YOLOX is one of the more practical options. It keeps the YOLO family's speed but removes anchor boxes, simplifying training without giving up accuracy. Pretrained weights run from a small model built for edge devices up to a full-size variant, so you can match model size to your hardware budget, and the COCO benchmarks spell out the speed and accuracy trade-off at each size. The repo also includes demos and export paths for ONNX, TensorRT, ncnn, and OpenVINO, which cuts a lot of the usual friction in moving a trained model to a real runtime. Teams that need to customize and ship a modern YOLO detector will find a dependable starting point here.