yolov7_d2

Why we included this project

YOLOv7_d2 is a detectron2-based training codebase that covers both classic YOLO detectors and transformer-based DETR models. You can train YOLOv4/YOLOv7, YOLOX, YOLOv6, and DETR-family architectures from one pipeline using COCO-format data and shared augmentation and loss utilities. It also goes beyond plain bounding boxes: instance segmentation and one-stage keypoint heads are supported, and many models export to ONNX and run with TensorRT, which matters if you plan to deploy a trained detector on edge hardware. The bundled backbones (ResNet, Res2Net, Swin, PVTv2, ConvNeXt) and interchangeable heads make it handy for measuring accuracy-versus-latency trade-offs on your own dataset before picking one architecture. Just be aware that some newer features are closed-source and the repo has not been updated since late 2023, so treat it as a research toolkit rather than a maintained production framework.