opencv: onnx 12: Inconsistent shape for ConcatLayer in function 'cv::dnn::ConcatLayerImpl::getMemoryShapes'

Ok so I exported both models with different opset, this is what I got

when giving size & opset  9 -> Upsample [X, scales]
when giving size & opset 10 -> Resize   [X, scales]
when giving size & opset 11 -> Resize   [X, roi, scales, sizes]

when giving scales & opset  9 -> Upsample [X, scales]
when giving scales & opset 10 -> Resize   [X, scales]
when giving scales & opset 11 -> Resize   [X, roi, scales]

So I could change my PR to this:

            int opset = model_proto.has_ir_version() ? (int)model_proto.ir_version() : -1;
            Mat scales;
            switch (opset)
            {
                case 12:
                case 11: scales = getBlob(node_proto, 2); break;
                default: scales = getBlob(node_proto, 1); break;
            }

@JulienMaille Please use this script used for generation of .onnx test files: https://github.com/opencv/opencv_extra/blob/4.5.0/testdata/dnn/onnx/generate_onnx_models.py

Yes, I think it’s easier to understand. But you need a change:

layerParams.set("width", shapes[3]);
layerParams.set("height", shapes[2]);

I confirm that editing my onnx and adding “zoom_factor_x”, “zoom_factor_y” will make it work. So what we would need is to read from scales input instead

EDIT: which seems to be done here: https://github.com/opencv/opencv/blob/199687a1c5b194481754d3efe14cb06a82efa805/modules/dnn/src/onnx/onnx_importer.cpp#L1689-L1700