LdDl
[FEATURE] Support for YOLO v9/v10
Do you plan to add support for v9 or v10?
Yeah, that is in the plans, but I'm not sure where to look for source code for postproccessing very last layer (YOLO). PRs are welcome
@LdDl Maybe in python implementation? And one more question is there any plan to support Detectron 2 by meta? It has free license and also can be useful.
@LdDl Maybe in python implementation? And one more question is there any plan to support Detectron 2 by meta? It has free license and also can be useful.
Is this Detectron2? https://github.com/facebookresearch/detectron2 I thought this has been abandoned since Nov 16, 2021
@LdDl I tried the coco trained version using rust ort and it is working very good. Not as performant like yolo but the accuracy is very good. I need an open licensed od model so I cant use yolo, thats why I tried this one. But there is the open-mmlab and it uses also this one I think.
lovely code in rust. It runs well on mac m1, throwing 2x faster than python. i will try on ubuntu Arm 25.05, it seems not be compatible runtime libs. thanks much indeed @LdDl
Yolov9 and v11 supported. Yolov10 is tough...YOLOv10 introduced an nms-free architecture where post-processing (top-k selection) is built into the model itself. Unfortunately, as I am aware of opencv-dnn does not support the top-k ONNX operation, causing broken inference results....May be in future I can figure it out or someone could patch v10 export to ONNX for opencv-dnn needs