I am trying to implement a custom object detection network without Tensorflow dependency. My weights are stored in nchw order in binary file and can easily be loaded into cv::Mat. However, I can't seem to find an example of how to build custom layers in dnn module. Official documentation is slightly confusing for me.