jhjin / caffe2-cpp

Image classification and detection C++ example using caffe2 library

Caffe2 C++ API example

The repository provides a basic image classification and detection example using Caffe2 shared library (.so). Tested on the Ubuntu 16.04 machine.

Dependencies

Install dependent packages via apt-get.

make dependency
make cudnn       # (optional) download publicly released cudnn v5

Download cudnn library and decompress under the ./cuda directory which will be used by Caffe2 library. For some reasons if you do not have an access to the latest cudnn, just proceed with make cudnn. It will download publicly available cudnn (v5) - quite ancient but still ok to experiment with.

./cuda/include/cudnn.h
./cuda/lib64/libcudnn.so.*

Before build Caffe2, update CMakeLists.txt according to your configuration if needed.

Build and run

The default make target will do all jobs for you - build caffe2 library, download a pretrained model (Squeeznet) and test images, compile and run the app. The example app is heavily based on Leo Vandriel's work.

make     # all at once

Or if you want to move slowly step-by-step

make model    # download model
make build    # build caffe2
make app      # compile app
make cnn      # run classification app (inference)
make rcnn     # run detection app (inference)

Desired outcome from the command is:

$ make cnn
==> init network
==> parse image list
==> prepare batch
==> feedforward
==> retrieve results
P( lemon | lemon.jpg ) = 0.949022
P( daisy | flower.jpg ) = 0.960899

$ make rcnn
==> using CPU
==> init network
==> parse image list
==> prepare batch (1 x 3 x 600 x 1116)
==> feedforward
==> retrieve results
P( person | street.jpg [185,186,85,260] ) = 0.999836
P( person | street.jpg [366,208,124,281] ) = 0.99977
P( person | street.jpg [754,191,108,178] ) = 0.992861
P( person | street.jpg [466,241,98,81] ) = 0.973806
P( person | street.jpg [914,180,88,219] ) = 0.95949
P( person | street.jpg [840,179,45,81] ) = 0.898059
P( person | street.jpg [921,177,82,72] ) = 0.849532
P( person | street.jpg [953,178,50,57] ) = 0.747317
P( person | street.jpg [464,245,107,186] ) = 0.702847
P( car | street.jpg [939,164,173,333] ) = 0.988054
P( car | street.jpg [556,198,30,24] ) = 0.834219
P( car | street.jpg [585,203,26,21] ) = 0.787161
P( car | street.jpg [515,199,37,24] ) = 0.756677
P( motorcycle | street.jpg [771,299,83,134] ) = 0.994598
P( truck | street.jpg [828,153,187,230] ) = 0.909398
P( truck | street.jpg [258,181,78,102] ) = 0.771324
P( couch | street.jpg [442,310,175,186] ) = 0.985259