towei/deepslidingshape

Deep Sliding Shapes for Amodal 3D Object Detection in RGB-D Images

Deep Sliding Shapes for Amodal 3D Object Detection in RGB-D Images

S. Song, and J. Xiao. (CVPR2016)

Compile code

Download CUDA 7.5 and cuDNN 3. You will need to register with NVIDIA.

cd code/marvin
./linux.sh

Prepare data

• download the processed RGBD data here by runing script:

  downloadData('../sunrgbd_dss_data','http://dss.cs.princeton.edu/Release/sunrgbd_dss_data/','.bin');

• or run dss_preparedata() to prepare your own data.

• download the image and hha images by runing script: downloadData('../image','http://dss.cs.princeton.edu/Release/image/','.tensor'); downloadData('../hha','http://dss.cs.princeton.edu/Release/hha/','.tensor');

3D region proposal network:

• You can download the precomputed region proposal for NYU and SUNRGBD dataset by runing script:

  downloadData('../proposal','http://dss.cs.princeton.edu/Release/result/proposal/RPN_NYU/','.mat');
  downloadData('../proposal','http://dss.cs.princeton.edu/Release/result/proposal/RPN_SUNRGBD/','.mat');

• To train 3D region proposal network and extract 3D region proposal cd code/matlab_code/slidingAnchor run dss_prepareAnchorbox() to prepare training data. run RPN_extract() to extract 3D region proposal. You may need the segmentation result here:

  downloadData('../seg','http://dss.cs.princeton.edu/Release/seg/','.mat');

• Pretrained model and network defination can be found here

3D object detection network:

1. change path in dss_initPath.m;
2. run dss_marvin_script(0,100,1,[] ,1,'RPN_NYU',1,[],0,0);
3. Pretrained model and network defination can be found here

Notes :

• If matlab system call fails, you can try to run the command directly.
• The rotation matrixes for some of the images in the dataset are different from the original SUNRGB-D dataset, so that the rotation only contains camera tilt angle (i.e. point cloud does not rotated on the x,y plane). We provide the data in this repo ./external/SUNRGBDtoolbox/Metadata/SUNRGBDMeta.mat. All the results and ground truth boxes provided in this repo are using this rotation matrix. To convert the rotation matrix you can reference the code "changeRoomR.m"