Different img size

[uneedcrydear]

Different img size

[UNeedCryDear]

Your model takes (1,3,1376,1376) input and its outputs are (1, 116487, 38) ，which means your onnx model isn't P6-onnx model.
So you should set YOLO_P6 false
https://github.com/UNeedCryDear/yolov5-seg-opencv-dnn-cpp/blob/7af65d92ed0629ebf0af8b6ae015b2578483fa5f/yolo_seg.h#L5

[mucahitrtn]

Thanks, it worked.

[lioneldaniel]

Do you have an idea for explaining why the bbox are visualy misplaced when exporting yolov5s-seg.pt to ONNX with --img-size set to either 2752 or 1376 (see images on the left), whereas the bbox are well placed when --img-size is set to 640 (see image on the right)?
Notice that the masks are always well placed.

These results are obtained on CPU after updating the following lines with the parameters shown is the picture above :

• yolo_seg.h:27
• yolov5_seg_utils.h:14

If yolov5_seg_utils.h:14 is not updated, an exception is raise in GetMask2 at yolov5_seg_utils.cpp:167

In addition, I have modified yolov5_seg_utils.cpp#L180 so that each instance has a different color.

[UNeedCryDear]

@lioneldaniel
WongKinYiu/yolov7#433
I have tried scaling before, but the results told me that only scaling in a direction smaller than the img size used for training is effective.

[lioneldaniel]

@UNeedCryDear it seems that scaling in exactly one direction works with a size bigger than the one used for training:

My configuration:

• execution configuration: OpenCV C++ >v4.6.0 (Oct 19, 2022)
• export configuration: YOLOv5 (docker image) v7.0-55-g632bf48 Python-3.8.10 torch-1.13.1+cu117 CPU + onnx 1.12.0 + onnx-simplifier 0.4.13, with these options:
  python export.py --weights $path/models/yolov5s-seg.pt --include onnx --simplify --img-size 1376 640 --device cpu --opset 12

[UNeedCryDear]

There will also be problems with scaling in one direction. At present, you can't see it by double scaling. If the multiple is larger, there will be problems.

In fact, it is the scaling problem of the model itself, not the problem caused by exporting .pt to .onnx. I can't explain why this is so. Maybe you need to go to the yolov5-official to ask questions.