In GSoC 2016, OpenCV introduced the new deep learning module dnn_modern under opencv_contrib. dnn_modern is running on the top of tiny-dnn, which is a high-level, header-only, and portable framework written in modern C++. Now we can load Caffe's pre-trained model via dnn_modern, but we want to make it more user-friendly, more efficient, and more flexible for all OpenCV users.

1. Tensorflow and Pytorch converters

• Continue TF parser https://github.com/tiny-dnn/tiny-dnn/pull/522
• Collaborate with https://github.com/szagoruyko/functional-zoo

1. Supporting GPU Training/Inference

• LibDNN https://github.com/naibaf7/libdnn
• CUDA/cuDNN

1. Providing pre-traind models

• Implement and provide pre-trained models such: inception, resnet, pix2pix, infogan, style transfer, stackgan...

1. IOS and ANDROID improved toolchain integration, examples, tutorials

Check the initial roadmap for v1.0.0a

Things TODO before the GSOC:

• Tensor implementation

  • Kernels
  • Layers
  • Computational Graph
  • Quantizzation?

• Check/Test Caffe converter with real examples