Recommended DNN Framework
Hi,
given want to use OpenCV's DNN module for several tasks in several projects: classification, detection, and segmentation. I know there are pre-trained models which I can download, but i will have to re-train for my application.
I have tried a few examples from the number model zoo and they worked.
I would now like to train a network from scratch or transfer train with my own classes, and load&run the resulting model in OpenCV.
I have tried this in Tensorflow but didnt succeed. I have failed so far to generate pb (and pbtxt)-Files which could be read by OpenCV. I followed tutorials for training and transfer training, preparation of deep learning frameworks out there, model Files, etc. But overal it was a rather un-satisfying experience.
My impression is that Tensorflow is great for researchers, and my focus is on application.
At least OpenCV supports a few of them (Torch; Caffe; Tensorflow; via ONNX: caffe2, mxnet and others)
I see many caffe models in is also about applying trained DNNs.
Now my question is: If I want to use an existing network configuration, just (transfer-)train it with my own data, and apply the examples, but caffe seems not to be maintained very well anymore?
Can you recommend any framework(s) for training that model in OpenCV, should play well with OpenCV? Because it is particularly well tested, because it has the shortest to-to list, because it has the most complete list of implemented layers, because the opencv team sees the future with a particular framework?I use a different DNN Framework from Tensorflow? And which one?
Much appreciated, thanks!
