// the find
open-mmlab/mmyolo
OpenMMLab YOLO series toolbox and benchmark. Implemented RTMDet, RTMDet-Rotated,YOLOv5, YOLOv6, YOLOv7, YOLOv8,YOLOX, PPYOLOE, etc.
MMYOLO is OpenMMLab's unified training framework for the YOLO family — YOLOv5 through v8, YOLOX, RTMDet, PPYOLOE, and rotated detection variants — all under one config-driven system. It's aimed at researchers who want to run fair apples-to-apples comparisons between detectors, and practitioners who want to fine-tune a YOLO variant on a custom dataset without rewriting boilerplate. RTMDet (their own contribution) hits 52.8 AP on COCO at 322 FPS on a 3090, which is genuinely good.
RTMDet is a real contribution, not just a wrapper — the technical report shows careful ablation of the CSPNeXt backbone and dynamic label assignment, and the numbers back it up. The modular backbone/neck/head registry means you can swap in timm backbones or mix components from MMDetection without forking the codebase. Rotated object detection support (RTMDet-R on DOTA) is rare in YOLO toolboxes and actually useful for aerial/satellite imagery work. Deployment configs for TensorRT, ONNX, and RKNN are included and tested, not just mentioned.
Last meaningful commit was July 2024 and the latest release (v0.6.0) is from August 2023 — YOLOv9, YOLOv10, and YOLO11 are all missing, so the 'benchmark' framing is already stale. The install chain (mim, mmengine, mmcv, mmdet, then mmyolo) is four deep and notoriously brittle — CUDA version mismatches between mmcv prebuilts and your torch install will eat an afternoon. The config system inherits MMDetection's Python-dict inheritance model, which is powerful but has a steep learning curve; debugging why a nested override didn't apply is non-obvious. No native video streaming or RTSP inference path — everything assumes batch image inputs, so real-time deployment requires writing your own loop.