danaroth/icvl
ICVL is a hyperspectral image dataset collected with a Specim PS Kappa DX4 hyperspectral camera and a rotating platform for spatial scanning. Currently the dataset contains 200 images and will be expanded gradually. Images have a spatial resolution of 1392 × 1300 and cover 519 spectral bands (400‑1000 nm, ~1.25 nm intervals). The dataset provides raw data in ENVI format and downsampled data in MAT format (31 spectral channels, 400‑700 nm, 10 nm intervals). The original dataset contains only clean images; a separate dataset for hyperspectral image denoising is taken from another paper.
Dataset description and usage context
Dataset Description
ICVL is a hyperspectral image dataset collected by the "Sparse Recovery of Hyperspectral Signal from Natural RGB Images" project. The data were acquired using a Specim PS Kappa DX4 hyperspectral camera and a rotating stage. Currently the collection includes 200 images and will continue to grow.
The acquisition resolution is 1392 × 1300, covering 519 spectral bands (400‑1000 nm, approximately every 1.25 nm). Raw files (.raw) are stored in ENVI format, with accompanying header (.hdr) files containing decoding information. Additionally, .mat files are provided, downsampled to 31 spectral channels from 400 nm to 700 nm at 10 nm intervals.
The original dataset consists solely of clean images. For hyperspectral image denoising benchmarks, the test set is taken from the paper "3D Quasi‑Recurrent Neural Network for Hyperspectral Image Denoising".
Quick Preview
Below are previews of a subset of images from the dataset:
 |  |  |  |
| 4cam_0411-1640-1 | 4cam_0411-1648 | bguCAMP_0514-1659 | bguCAMP_0514-1711 |
| ... (additional image previews omitted for brevity) ... |
Pair the dataset with AI analysis and content workflows.
Once the source passes your review, move straight into summarization, transformation, report drafting, or presentation generation with the JuheAI toolchain.