README.md

IPS Generator

IPS Generator is a toolkit for generating the synthetic polarimetric stereo dataset named as IPS dataset in our paper ( DPS-Net: Deep Polarimetric Stereo Depth Estimation ).

1. Introduction

We present IPS Generator to synthesize the polarimetric data of IPS from the accurate normal map provided by the IRS. The surface normal is decomposed as the azimuth angle and the zenith angle to calculate the AoLP image and the DoLP image dominated by the specular reflection or diffuse reflection. The dominant reflection types are determined by the segmentation results of the SeMask-Segmentation, which was SOTA during paper submission. SAM is recommended for better segmentation quality. The synthetic DoLP and AoLP maps in the IPS dataset are finally obtained after adding Gaussian noise into the polarized images.

2. Installation

2.1. Python Environment

For the python environment, please refer to SeMask-Segmentation.

Then, init SeMask submodule with

git submodule --init --recursive

Next, the pre-trained model semask_large_mask2former_ade20k.pth for segmentation can be download, following SAM.

2.2. C++ Environment

First, please install the following dependencies

• Eigen
• OpenCV
• PCL

Then, the cpp code can be compiled

mkdir build
cd build
cmake ..
make -j

3. Usage

First, run following python scripts to segment the RGB images for generating the dominant reflection type

python sem_seg.py

Then, synthesis DoLP and AoLP can be produced by executing the executable file.

./rebuild_by_ratio_distribution

4. Real Data

The real polarimetric dataset is captured as well. The RPS dataset utilized in DPS-Net can be download from Google Drive.