Hyper-Object Challenge @ ICASSP 2026

Reconstructing Hyperspectral Cubes of Everyday Objects from Low-Cost Inputs

Learn More Github

About the Challenge

Hyperspectral imaging (HSI) captures fine-grained spectral information, enabling precise material analysis. However, commercial HSI cameras are expensive and bulky. In contrast, RGB cameras are affordable and ubiquitous.

The Hyper-Object Challenge aims to revolutionize access to spectral imaging by developing learning-based models that can reconstruct high-fidelity hyperspectral cubes (400-1000 nm) from low-cost inputs like mosaic sensor data or standard low-resolution RGB images. This would unlock applications from counterfeit detection to agricultural monitoring using everyday devices.

Example images from the Hyper-Object dataset

Diverse everyday objects from the Hyper-Object dataset.

Challenge Tracks

The competition comprises two distinct tracks, each defined by a different low-cost input format i_k.

Track 1: Spectral Reconstruction from Mosaic Images

Each input i_k is a mosaic image that captures only one spectral band per pixel.

\( i_k \in \mathbb{R}^{H \times W \times 1} \)

Every pixel is filtered according to a predefined tiling pattern that mimics a snapshot spectral sensor. The task is to reconstruct the full spectral cube

\( h_k \in \mathbb{R}^{H \times W \times C} \)

Objective

Recover all C = 61 spectral bands (400 – 1000 nm, 10 nm spacing) from the spectrally-subsampled observation.

Evaluation

Submissions are ranked by Spectral-Spatial-Color (SSC) score, range in [0, 1].

Spectral: SAM, SID, ERGAS
Spatial: PSNR, SSIM on a standardized sRGB render (D65, CIE 1931 2°)
Color: ΔE00 on the same sRGB render

Track 2: Joint Spatial and Spectral Super-Resolution

Each input i_k is a low-resolution RGB image captured by a commodity camera.

\( i_k \in \mathbb{R}^{h \times w \times 3}, \qquad h \ll H,\; w \ll W \)

The goal is to jointly recover spatial and spectral resolution by reconstructing

\( h_k \in \mathbb{R}^{H \times W \times C} \)

where C ≫ 3 and H ≫ h, W ≫ w.

Objective

Produce a high-fidelity hyperspectral cube that restores full spatial and spectral resolution.

Evaluation

Submissions are ranked by Spectral-Spatial-Color (SSC) score, range in [0, 1].

Spectral: SAM, SID, ERGAS
Spatial: PSNR, SSIM on a standardized sRGB render (D65, CIE 1931 2°)
Color: ΔE00 on the same sRGB render

Important Dates (Tentative)

20 August 2025

Registration Opens

25 August 2025

Public Testing Set & Baseline Release

31 August 2025

Private Testing Set Release & Leaderboard Starts

25 November 2025

Submission Closes

30 November 2025

Winner Announcements

07 December 2025

2-page Papers Due (by invitation)

11 January 2026

Paper Acceptance Notification

18 January 2026

Camera-ready Papers Due

Submissions will only be considered if accompanied by the full code.
Detailed instructions for code submission will be provided one month before the submission deadline.

The top five winners will be selected based on code evaluation on the hidden test set.
Participants must ensure that their code runs successfully on our end for the final evaluation.

Winners will be notified via email.

How to Participate

Access Resources

Get the training data, baseline models, and evaluation scripts on our Kaggle page.

Develop Your Model

Train your learning-based model to reconstruct high-fidelity hyperspectral cubes.

Submit & Compete

Submit your predictions on the held-out test set and climb the leaderboard.

The competition will be hosted on Kaggle. Click below to go to the competition page, register your team, and get started!

Organizers

Pai Chet Ng

Singapore Institute of Technology

Konstantinos N. Plataniotis

University of Toronto

Juwei Lu

University of Toronto

Gabriel Lee Jun Rong

Singapore Institute of Technology

Malcolm Low

Singapore Institute of Technology

Nikolaos Boulgouris

Brunel University

Thirimachos Bourlai

University of Georgia

Seyed Mohammad Sheikholeslami

University of Toronto

Leaderboard

Baseline results on the public validation set.
Final ranking will use the composite score that combines RMSE, SAM, PSNR, SSIM, and EGRAS on the hidden test set.

Track 1 Track 2

Rank	Team	Composite Score	Submissions	Last Update
1 🥇	Baseline Model (MST++)	0.726	Baseline	-
2 🥈	Baseline Model (HRNet)	0.560	Baseline	-
3 🥉	Baseline Model (HSCNN+)	0.350	Baseline	-
4	Your Team	TBD	—	—

Rank	Team	Composite Score	Submissions	Last Update
1 🥇	Baseline Model (MST++)	0.705	Baseline	-
2 🥈	Baseline Model (HRNet)	0.533	Baseline	-
3 🥉	Baseline Model (HSCNN+)	0.310	Baseline	-
4	Your Team	TBD	—	—

Frequently Asked Questions

What is the goal of the Hyper-Object Challenge?

The Hyper-Object Challenge pushes the frontier of low-cost hyperspectral imaging. Participants are invited to develop deep learning models that reconstruct high-fidelity hyperspectral image cubes (400–1000 nm, 61 bands) of real-world objects from cost-effective input formats, such as spectral mosaics or low-resolution RGB images. The goal is to democratize access to spectral imaging technology for real-world applications like agriculture, cultural heritage, forensics, and material analysis.

What are the competition tracks?

We offer two tracks, each focusing on a different practical constraint:

Track 1 – Spectral Reconstruction from Mosaic Images
Input: A single-channel spectral mosaic image (1 band per pixel)
Output: A full-resolution hyperspectral cube with 61 bands
Track 2 – Joint Spatial & Spectral Super-Resolution
Input: A low-resolution RGB image (captured by a commodity camera)
Output: A high-resolution hyperspectral cube (61 bands, upsampled)

You may participate in one or both tracks.

What data is provided?

Training data – Available now
Public test data – Released on Aug 25, includes input and ground-truth for local validation
Private test data – Input-only; used for leaderboard ranking
Hidden test data – Used internally to evaluate your submitted model/code for final ranking

Please refer to the Data Description section for more details.

How is scoring done?

We use the Spectral-Spatial-Color (SSC) composite score as the main evaluation metric. It combines the following metrics:

SAM (Spectral Angle Mapper)
SID (Spectral Information Divergence)
ERGAS
PSNR
SSIM
ΔE (color difference in rendered RGB space)

Each metric is normalized and combined into a single SSC score in [0, 1], where higher is better.

What should my model output?

Each sample's output should be a 61-band hyperspectral image cube, matching the spatial resolution of the input (Track 1) or a defined high-res target (Track 2). Outputs must be saved in .npy format as *.npy files named after the test sample ID.

Can I use external datasets or pre-trained models?

Yes. External datasets, synthetic data, and pre-trained models are allowed, but must be clearly disclosed in your final write-up. Please do not use test set leakage or non-public data.

Note: The organizers and affiliated students will not participate in the competition.

What is the submission format?

You must submit a .zip file containing one .npy file per test sample, named exactly as per the sample ID (e.g., sample123.npy).
The output of each .npy file must be a float32 H×W×61 array in the range [0, 1].

What is the leaderboard setup?

The leaderboard reflects scores on the private test set.
Your final ranking will be based on internal evaluation on the hidden test set using your submitted code.
This ensures fair and robust evaluation under realistic deployment conditions.

How do I submit my model for final evaluation?

In addition to .zip file predictions, you must submit:

Your trained model weights
A standalone inference script (e.g., inference.py) that loads your model and generates output .npy files for any given input directory
A requirements.txt listing all dependencies

We will run your code in a standardized environment and evaluate on the hidden test set.

What is the submission limit?

You may make up to 5 submissions per day
A maximum of 3 final code submissions can be made for internal evaluation
Your highest SSC score from those 3 code submissions will be used for the final ranking

What are the hardware or runtime constraints?

For the final evaluation, your model must be runnable on a single GPU (e.g., NVIDIA RTX 6000 Ada or similar) with <24GB VRAM and <30 minutes inference time per test sample. If your model exceeds this, you may be disqualified.

Where can I ask questions or get help?

Please post your questions on the Kaggle discussion forum. The organizers will actively respond and clarify rules and requirements.

What license applies to the data?

The dataset is released under the MIT License, which permits free use, modification, and distribution for both research and commercial purposes, provided that the original authors are properly credited.

What do the winners receive?

Featured in a special session at ICASSP 2026
Invited to contribute to the joint overview paper with the organizers
Provided with certificates and digital badges