Skip to content

Latest commit

 

History

History
executable file
·
49 lines (39 loc) · 2.55 KB

README.md

File metadata and controls

executable file
·
49 lines (39 loc) · 2.55 KB

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

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"