TensorFlow C++ inference of FCRN-Depth-Prediction

This is a TensorFlow c++ inference code of FCRN-DepthPrediction (single-view deep depth prediction) for indoor scenes.

This code uses cppflow, which doesn't require either TensorFlow compile or bazel (this installation is known as a very difficult task [1, 2]).

Environment

• Ubuntu18.04 (GPU: NVIDIA GeForce GTX 1080)
• CUDA10.0
• cuDNN7.4
• Python3.6.9 (for preparing pb-file)
• TensorFlow for C 1.13.1 (this version requires CUDA10.0 & cuDNN7.4)

You need to download libtensorflow & install it referring to official page.

wget https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-gpu-linux-x86_64-1.13.1.tar.gz
sudo tar -C /usr/local -xzf libtensorflow-gpu-linux-x86_64-1.13.1.tar.gz
sudo ldconfig

How to run it

See README in freeze_graph, and prepare pb-file of pretrained FCRN-DepthPrediction model.

cd freeze_graph/pb_file
./download.sh

Build our code (based on cppflow) with following commands.

mkdir build && cd build
cmake ..
make

Specifying the path to the pb-file & target image directory (samples from NYUDepthv2), execute as follows.

./inference_fcrn \
    pb_file="../freeze_graph/pb_file/NYU_FCRN.pb" \
    img_dir="../samples/home_office_0013"

References

• FCRN-DepthPrediction
• cppflow
• monodepth-cpp
• NYU Depth Dataset v2