im2pano3d

im2pano3d copied to clipboard

Im2Pano3D: Extrapolating 360° Structure and Semantics Beyond the Field of View

This repo contains training and testing code for our paper on Im2Pano3D. More information about the project can be found in our paper and project webset

Organization

The code and data are organized as follows:

    im2pano3d
         |-- matlab_code
         |-- torch_code
         |-- models
         |-- datalist
         |-- data
                |-- mpv3  
                |-- suncgpano  

Installation

0. Requirements for torch (see: torch installation instructions)
1. Install torch-hdf5

sudo apt-get install libhdf5-serial-dev hdf5-tools
git clone https://github.com/deepmind/torch-hdf5
cd torch-hdf5
luarocks make hdf5-0-0.rockspec LIBHDF5_LIBDIR="/usr/lib/x86_64-linux-gnu/"

Download

0. Dwonload training and testing panorama data: ./download_pano_data.sh [41 G]
1. Dwonload models: ./download_models.sh [1.7 G]

Training

Export path

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cudnn/v5/lib64

Train on SUNCG data

cd torch_code/
name=suncg maskType=twoview   loss_xyz=1 loadOpt=rgbpns  Gtype_in=rgbpns   Gtype_out=pns  dataset=suncg  dataPath=../data/suncgpano/ DATA_ROOT=../datalist/trainlist_suncgroom8_10000.txt gpu=1 th train.lua  2>&1 | tee ./checkpoints/training_suncg.log

Train on Matterport3D data

cd torch_code/
name=mp maskType=twoview   loss_xyz=1 loadOpt=rgbpns  Gtype_in=rgbpns   Gtype_out=pns  dataset=mp  dataPath=../data/mpv3/  DATA_ROOT=../datalist/trainlist_mp3.txt gpu=1 th train.lua  2>&1 | tee ./checkpoints/training_suncg.log

0. Supported input output modalities:

0. Supported mask types (maskType):
   twoview: Input two views of the panorama and predicts the other two. (a)
   1camera: one normal camera (b)
   3camera: three normal cameras (c)
   middlecamera: Input six RGB-D camera looking horizantoally forward (d)
   upcamera: Input six RGB-D camera looking upwards (e)
   random: random mask
   nomask: No mask applied in the input image. Example usage, input rgb panorama output pns.(f)\

Testing

name=mp_twoview_rgbpns  maskType=twoview dataset=suncg dataPath=../data/suncgpano/ checkpoints_dir=../checkpoints  DATA_ROOT=../datalist/testlist_suncgroom8.txt how_many=480 th test.lua

Testing with pre-trained models:

name=mpft_suncg_pns_pns_twoview  maskType=twoview dataset=suncg dataPath=../data/suncgpano/  checkpoints_dir=../models DATA_ROOT=../datalist/testlist_suncgroom8.txt how_many=480 th test.lua

Depth encoding conversion

See ./matlab_code/pano_io/pano_demo.m. It shows how to \

1. Compute the 3D pointcloud from depth representation. \
2. Compute the 3D pointcloud from plane equation representation. \

Acknowledgment

Code is inspired by DCGAN, DCGAN, Context-Encoder and pix2pix.