Different results converting blob from dnn module to cv::Mat in release and debug mode


After converting an onnx network to dnn with openvino, converting the output blob to cv::Mat works just fine in the VIsual Studio 2019 debug mode. But when switching to release mode only one coordinate (always the same, the y2 coordinate of all predicted boxes), is randomly false. In all release runs it has the same false values.

Is there anything about the following conversion that could produce this error:

	cv::Mat frame_normed;
		frame_resized.convertTo(frame_normed, CV_32F, 1.f / 255);
		//cv::Mat blob;
		//// reshape to reach model inputs
		frame_normed = frame_normed.reshape(0, model.getInputShape());

		// set model inputs
		model.getNetwork_().setInput(frame_normed, "input");

		// execute model and convert output to cv::Mat
		auto output_blob = model.getNetwork_().forward();

		int output_dims[2] = { config.output_shape[0], config.output_shape[1] };
		cv::Mat predicted_output = cv::Mat(output_dims[0], output_dims[1], CV_32F, output_blob.ptr<float>());

Also it seems weird that the dimensions of frame_normed and output_blob are just like 1x-1x-1

Best regards