Created by Yu Xiang and Tanner Schmidt at RSE-Lab at University of Washington.

We introduce Data Associated Recurrent Neural Networks (DA-RNNs), a novel framework for joint 3D scene mapping and semantic labeling. DA-RNNs use a new recurrent neural network architecture for semantic labeling on RGB-D videos. The output of the network is integrated with mapping techniques such as KinectFusion in order to inject semantic information into the reconstructed 3D scene. arXiv, Video

DA-RNN is released under the MIT License (refer to the LICENSE file for details).

If you find DA-RNN useful in your research, please consider citing:

@inproceedings{xiang2017darnn,
    Author = {Yu Xiang and Dieter Fox},
    Title = {DA-RNN: Semantic Mapping with Data Associated Recurrent Neural Networks},
    Booktitle = {Robotics: Science and Systems (RSS)},
    Year = {2017}
}

DA-RNN consists a reccurent neural network for semantic labeling on RGB-D videos and the KinectFusion module for 3D reconstruction. The RNN and KinectFusion communicate via a Python interface.

1. Install TensorFlow. I suggest to use the Virtualenv installation.

2. Compile the new layers under $ROOT/lib we introduce in DA-RNN.

   cd $ROOT/lib
   sh make.sh

3. Compile KinectFusion with cmake. Unfortunately, this step requires some effort.

   Install dependencies of KinectFusion:

   • Pangolin
   • Eigen
   • Sophus
   • nanoflann
   • libsuitesparse-dev

   cd $ROOT/lib/kinect_fusion
   mkdir build
   cd build
   cmake ..
   make

4. Compile the Cython interface for RNN and KinectFusion

   cd $ROOT/lib
   python setup.py build_ext --inplace

5. Add the KinectFusion libary path

   export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$ROOT/lib/kinect_fusion/build

6. Download the VGG16 weights from here (57M). Put the weight file vgg16_convs.npy to $ROOT/data/imagenet_models.

• Ubuntu 16.04
• Tensorflow 1.2.0
• CUDA 8.0

1. Download the RGB-D Scene dataset from here (5.5G).

2. Create a symlink for the RGB-D Scene dataset

   cd $ROOT/data/RGBDScene
   ln -s $RGB-D_scene_data data

3. Training and testing on the RGB-D Scene dataset

   cd $ROOT
   
   # train and test RNN with different input (color, depth, normal and rgbd)
   ./experiments/scripts/rgbd_scene_multi_*.sh $GPU_ID
   
   # train and test FCN with different input (color, depth, normal and rgbd)
   ./experiments/scripts/rgbd_scene_single_*.sh $GPU_ID
   

1. Download the ShapeNet Scene dataset from here (2.3G).

2. Create a symlink for the ShapeNet Scene dataset

   cd $ROOT/data/ShapeNetScene
   ln -s $ShapeNet_scene_data data

3. Training and testing on the RGB-D Scene dataset

   cd $ROOT
   
   # train and test RNN with different input (color, depth, normal and rgbd)
   ./experiments/scripts/shapenet_scene_multi_*.sh $GPU_ID
   
   # train and test FCN with different input (color, depth, normal and rgbd)
   ./experiments/scripts/shapenet_scene_single_*.sh $GPU_ID
   

1. You can download all our trained tensorflow models on the RGB-D Scene dataset and the ShapeNet Scene dataset from here (3.1G).

   # an exmaple to test the trained model
   ./experiments/scripts/rgbd_scene_multi_rgbd_test.sh $GPU_ID