// the find
Megvii-BaseDetection/YOLOX
YOLOX is a high-performance anchor-free YOLO, exceeding yolov3~v5 with MegEngine, ONNX, TensorRT, ncnn, and OpenVINO supported. Documentation: https://yolox.readthedocs.io/
YOLOX is an anchor-free object detector from Megvii that outperformed YOLOv3-v5 when it dropped in 2021. It introduced decoupled heads and SimOTA label assignment to the YOLO family, and ships pretrained weights across a full size ladder from Nano (0.9M params) to X (99M params). It's for practitioners who need a well-benchmarked detector with serious deployment coverage: TensorRT, ONNX, ncnn, OpenVINO, and MegEngine all have working demos.
The decoupled head design — separating classification and regression branches — was a meaningful architectural improvement over the coupled heads in YOLOv3-v5 and the benchmark numbers back it up. Deployment story is unusually thorough: C++ and Python examples for TensorRT, ncnn, OpenVINO, and ONNX, plus an Android demo that actually runs. SimOTA label assignment is well-implemented and the assignment visualization tool added in 2023 makes it debuggable, which most detection repos skip entirely. The exp-file config system is cleaner than most YOLO repos — each variant is a Python class that overrides a base, so you can customize training without touching core code.
The repo is effectively in maintenance mode — last meaningful update was 2022, and the object detection field has moved substantially since then (RT-DETR, YOLOv8, YOLO11 all exist now and are actively developed). The benchmark numbers are measured on V100 which nobody buys anymore; there are no A100/H100 or edge-device latency numbers. COCO and VOC are the only supported dataset formats out of the box — getting custom data in requires reading a tutorial and writing glue code, which is more friction than YOLOv8's one-liner. The test suite is nearly empty: one file testing model utilities, nothing covering data augmentation, loss functions, or the NMS implementation.