System information (version)

• OpenCV => 4.3
• Operating System / Platform =>Windows
• Compiler => MSVC 16 (VS 2019 Community)

Detailed description

Basically I am trying to run deeplab with mobilnet backbone in dnn on cpu.
I managed to get good results in Python with Tensorflow.
At first I got errors due to not supporting some end layers, but I removed them (basically final argmax) so I got output as 256x256x2 image where after computing argmax by hand between those two channels I should be able to obtain binary mask.

However I get cv::Mat of correct size but whenever I access float values they make no sense
e.g. 361197504.0 where in baseline implementation are like 2.7776315 (min: -5, max: 5).
What's more, the values are not changing much when I change images which makes me think OpenCV returns some junk memory without raising exception.
The final result comes from layer of type ResizeBilinear.

Steps to reproduce

I am attaching the network intended for people semantic segmentation (input is 256x256x3, colors normalized to range (-1,1) ).

dl256.zip

I guess code used to test would be:

auto input_image = cv::imread("a.png");
auto  net = cv::dnn::readNetFromTensorflow("dl256.pb");
net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);
auto output_names = GetOutputsNames(net);
cv::Mat blob;
cv::dnn::blobFromImage(input_image, blob, 2 / 255.0, cv::Size(256, 256), cv::Scalar(127.5, 127.5, 127.5), true, false);
net.setInput(blob);
auto invalid_values = net.forward(output_names.front()); // there is one output layer

Expected: values in range (-5, 5), values changes for different images or at least exception that inference failed.
Example image:

Issue submission checklist

• I report the issue, it's not a question
• I checked the problem with documentation, FAQ, open issues,
  answers.opencv.org, Stack Overflow, etc and have not found solution
• I updated to latest OpenCV version and the issue is still there
• There is reproducer code and related data files: videos, images, onnx, etc