OpenCV 3.3のDNN - Vengineerの妄想

Vengineerの戯言 : Twitter、Slideshare
SystemVerilogの世界へようこそ、すべては、SystemC v0.9公開から始まった

　The main news is that we promoted DNN module from opencv_contrib to the main repository, 
　improved and accelerated it a lot. An external BLAS implementation is not needed anymore. 

　For GPU there is experimental DNN acceleration using Halide (http://halide-lang.org). 
　The detailed information about the module can be found in our wiki: Deep Learning in OpenCV.

　OpenCV can now be built as C++ 11 library using the flag ENABLE_CXX11. 
　Some cool features for C++ 11 programmers have been added.

　We've also enabled quite a few AVX/AVX2 and SSE4.x optimizations in the default build of OpenCV 
　thanks to the feature called 'dynamic dispatching'. 
　The DNN module also has some AVX/AVX2 optimizations.

　Intel Media SDK can now be utilized by our videoio module to do hardware-accelerated video encoding/decoding. 
　MPEG1/2, as well as H.264 are supported.

　Embedded into OpenCV Intel IPP subset has been upgraded from 2015.12 to 2017.2 version, 
　resulting in ~15% speed improvement in our core & imgproc perf tests.

　・DNNモジュールがcontribからmainに昇格！外部のBLAS無しでも性能改善も。
　・GPU利用時は、Halideを使ってDNNをアクセラレート。。。え、GPUだけなの？

Deep Learning in OpenCVによると、
サポートするフレームワークは、Caffe 1、TensorFlow、Torch/PyTorch。
Chainerはサポートしていないのね。

Halide利用は、OpenCL時なので、CUDAではないのね。

これによると、CPUでもHalide使えるのね。。。
でも、特別速いわけではssor (RoCC) accelerators for Rocket Chip]

追記)、2017.08.12
OpenCV 3.3.0で dnn のサンプルを試してみた。 (I tried dnn sample program with OpenCV 3.3.0)
OpenCV 3.3.0で dnn のサンプル (semantic segmentation) を試してみた。その２　(I tried semantic segmentation sample with OpenCV 3.3.0. part2)
OpenCV 3.3.0で dnn のサンプル (SSD) を試してみた。その３　(I tried Single Shot MultiBox Detector sample with OpenCV 3.3.0. part3)