This repository contains supporting code for the NTIRE 2020 Spectral Reconstruction challenge held in conjunction with CVPR 2020.
The challenge includes two tracks:
-
Track 1: “Clean” recovering hyperspectral data from uncompressed 8-bit RGB images created by applying a known response function to ground truth hyperspectral information.
-
Track 2: “Real World” recovering hyperspectral data from jpg-compressed 8-bit RGB images created by applying an unknown response function to ground truth hyperspectral information.
450 hyperspectral training images and their corresponding "Clean" and "Real World" images are available on the challenge track websites above, registration is required to access data.
The clean_example.ipynb
and real_world_example.ipynb
Jupyter notebooks include example code demonstrating how "Clean" and "Real World" training images were created.
SpectralUtils.py
includes utilities for handling spectral images and projecting them to RGB.
EvalMetrics.py
includes code used to measure reconstruction accuracy - there are the metrics which shall be used to score participants during the challehnge.
The resource directory contains:
cie_1964_w_gain
- the response function used in the "Clean" track.example_D40_camera_w_gain
- an example physical camera response function, somewhat similar to that used in the "Real World" track.sample_hs_img_001.mat
- a sample spectral image, additional images are available on the challenge track websites above.
While SpectralUtils
contains some example noise parameters, and an example camera response function is included in the resources
folder, the "Real World" track will be using different noise parameters and a different camera response function which will remain confidential throughout the challenge.