README.md

metadata

license: cc-by-nc-nd-4.0
task_categories:
  - image-to-image
  - image-feature-extraction
tags:
  - hyperspectral
  - spectral-imaging
  - remote-sensing
  - computer-vision
  - natural-images
  - icvl
  - bgu
pretty_name: ICVL Hyperspectral Dataset (2016)
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: preview/**

ICVL Hyperspectral Dataset (2016)

202 natural-scene hyperspectral images · 400–1000 nm · Specim PS Kappa DX4
_{Interdisciplinary Computational Vision Laboratory · Ben-Gurion University of the Negev}

🌐 Project page · 📄 Citation · ⚖️ License

---

🔭 Overview

The ICVL Hyperspectral Dataset is a collection of high-resolution hyperspectral images of natural scenes — urban landscapes, rural views, indoor environments, plants, everyday objects and more — released by the Interdisciplinary Computational Vision Laboratory (ICVL) at Ben-Gurion University of the Negev.

Images were acquired with a Specim PS Kappa DX4 hyperspectral camera mounted on a rotary stage for spatial (line) scanning, yielding dense spectral cubes across the visible and near-infrared range.

📊 Specifications

Property	Raw data (ENVI)	MAT (downsampled)
Spectral range	400 – 1000 nm	400 – 700 nm
Spectral bands	519 (~1.25 nm step)	31 (10 nm step)
Spatial resolution	1392 × 1300	1392 × 1300
Format	.raw + .hdr	.mat (HDF5)
Scenes	202	202

📁 Repository structure

ICVL_HS_2016/
├── mat/          # 202 files — downsampled 31-band cubes (400–700 nm)
│   └── <scene>.mat
├── raw/          # 404 files — full 519-band ENVI cubes (400–1000 nm)
│   ├── <scene>.raw
│   └── <scene>.hdr
├── preview/      # 202 files — RGB JPEG previews
│   └── <scene>.jpg
├── file_list.txt # optional splits / subset list
└── README.md

Total: 809 data files · ~335 GB.

Scene stems (e.g. 4cam_0411-1640, bguCAMP_0514-1711) are shared across mat/, raw/ and preview/, so each scene can be referenced by its stem.

🖼️ Sample scenes

_{RGB previews of hyperspectral cubes — plants, urban scenes, indoor objects.}

🚀 Loading the data

Downsampled .mat cubes (recommended for most vision tasks)

MATLAB v7.3 (HDF5-backed). Each file contains:

• rad — hyperspectral radiance cube, shape (1300, 1392, 31), wavelengths 400 – 700 nm at 10 nm steps.
• bands — vector of 31 wavelength values (nm).

import h5py, numpy as np
with h5py.File("mat/<scene>.mat", "r") as f:
    cube  = np.array(f["rad"])   # transpose to (H, W, C) as needed
    bands = np.array(f["bands"])

Full-range ENVI .raw + .hdr cubes

519 spectral bands over 400 – 1000 nm. Read with the spectral Python package or any ENVI-aware tool:

import spectral
img  = spectral.envi.open("raw/<scene>.hdr", "raw/<scene>.raw")
cube = img.load()                # (1300, 1392, 519)

RGB previews

Standard sRGB JPEGs for quick browsing and dataset navigation — no radiometric use.

🎯 Intended uses

• Spectral super-resolution — recovering hyperspectral signals from RGB inputs
• Hyperspectral denoising, super-resolution, compression
• Spectral unmixing and material identification
• Benchmarking spectral–spatial deep-learning models (NTIRE spectral-recovery challenges have used this dataset extensively)

⚖️ License

Released under CC BY-NC-ND 4.0 — Attribution · NonCommercial · NoDerivatives.

You may share the dataset for non-commercial purposes with attribution to the original authors; derivative distributions are not permitted.

📄 Citation

If you use this dataset in academic work, please cite:

@inproceedings{arad_and_ben_shahar_2016_ECCV,
  title        = {Sparse Recovery of Hyperspectral Signal from Natural RGB Images},
  author       = {Arad, Boaz and Ben-Shahar, Ohad},
  booktitle    = {European Conference on Computer Vision (ECCV)},
  pages        = {19--34},
  year         = {2016},
  organization = {Springer}
}

📮 Contact

Interdisciplinary Computational Vision Laboratory (ICVL) Ben-Gurion University of the Negev 🔗 icvl.cs.bgu.ac.il — Hyperspectral Imaging

Correspondence: Prof. Ohad Ben-Shahar (PI) — ben-shahar@cs.bgu.ac.il