Deep Hyperspectral-Depth Reconstruction Using Single Color-Dot Projection

This is the official repo for the implementation of the dataset generateion part of the CVPR2022 paper: Deep Hyperspectral-Depth Reconstruction Using Single Color-Dot Projection.

Introduction

Since it is difficult to simultaneously acquire accurate depth and spectral reflectance as a large-scale ground-truth dataset in real-world situations, we developed a spectral renderer to generate a synthetic dataset with rendered RGB color-dot images, ground-truth disparity maps, and ground-truth spectral reflectance images by extending the algorithm of a structured-light renderer.

Usage

Dependencies

The python packages can be installed with anaconda:

conda install --file requirements.txt

Building

First make sure the correct CUDA_LIBRARY_PATH is set in config.json. Afterwards, the renderer can be build by running make within the renderer directory.

Running

First, download ShapeNet V2 and change SHAPENET_ROOT in config.json. Then the data can be generated and saved to DATA_ROOT in config.json by running

python create_syn_data.py

Example

The generated dataset example including 8448 scenes for training (458.6GB) and 256 scenes for testing (13.9GB) can be downloaded here.

Acknowledgement

The code structure and some code snippets (rasterisation, shading, etc.) are borrowed from Connecting the Dots. Thanks for this great project.