// the find
lucasjinreal/yolov7_d2
🔥🔥🔥🔥 (Earlier YOLOv7 not official one) YOLO with Transformers and Instance Segmentation, with TensorRT acceleration! 🔥🔥🔥
A detectron2-based object detection framework that wraps YOLOX, YOLOv6, DETR, AnchorDETR, SparseInst, and several other architectures under one training harness. It predates the official WongKinYiu YOLOv7 and the name overlap is a source of genuine confusion. Aimed at researchers who want to mix-and-match backbones and heads without rewriting training loops.
The backbone zoo is unusually wide — Swin, PVTv2, ConvNeXt, Res2Net, EfficientNet, MobileViT all wired into detectron2's registry, so swapping them is a config change. ONNX and TensorRT export paths exist for DETR and SparseInst, including int8 quantization scripts, which most academic repos skip entirely. SparseInst integration is well-done: pretrained weights ship with it, export works, and there is a working int8 quantized ONNX that runs on CPU without unsupported ops. The config-driven multi-head design makes it straightforward to bolt a keypoint or segmentation head onto a detection backbone without touching model code.
The repo has not been touched since late 2023 and targets torch 1.11, which is three major versions behind; ONNX export in particular has known version sensitivity and will likely break on current toolchains. Key features are explicitly paywalled behind manaai.cn — the README lists ConvNeXt training, GFL loss, MobileViT-V2, and CSPRep-Resnet as closed-source, so what you get here is deliberately incomplete. The naming collision with the official WongKinYiu YOLOv7 causes real friction: search results, issues, and blog posts bleed between the two projects. Performance tables in the README are mostly marked 'tbd', so you cannot benchmark architectures against each other without running your own experiments.