# 2026 ICASSP Hyper-Object Challenge: Track 2 — Data

来源：https://www.kaggle.com/competitions/2026-icassp-hyper-object-challenge-track-2/data

本地 Meta Kaggle 快照；不是本次实时页面。

# Dataset Description

The **Hyper-Object Challenge (Track 2)** provides multiple datasets split into training and testing partitions. The competition focuses on joint spatial and spectral super-resolution from low-resolution RGB inputs.

## Dataset Splits

- **Training Data (Available Now):**  
  - Contains paired inputs (low-resolution RGB) and ground truth hyperspectral (HSI) cubes.  
  - Total: **167 training samples**.

- **Public Testing Data (Available by Aug 25):**  
  - Will be released with inputs only (low-resolution RGB).  
  - Used for participants to test their models locally.

- **Private Testing Data:**  
  - Contains inputs (low-resolution RGB).  
  - Ground truth will not be provided.  
  - Used for **Kaggle leaderboard scoring**.

- **Hidden Testing Data:**  
  - Reserved internally for final evaluation.  
  - Participants’ submitted inference code will be executed on this hidden dataset to determine the official final rankings.

---
## Folder Layout (Track 2)
<pre style="font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, 'Liberation Mono', monospace; line-height: 1.35;">
2026-Hyper-Object-Data/
|-- train/
|   |-- rgb_2/        # {id}.png   (h, w, 3)
|   `-- hsi_61/        # {id}.h5    (H,W,61)    dataset "cube", float32 in [0,1]
|-- test_public/
|   |-- rgb_2/        # {id}.png   (h, w, 3)
|   `-- hsi_61/        # {id}.h5    (H,W,61)    dataset "cube", float32 in [0,1]
`-- test_private/
    `-- rgb_2/        # {id}.png
</pre>
---

## Training Data Details

### 1. Low-Resolution RGB (LR-RGB)
- Generated from each ground-truth HSI cube by first rendering RGB with **CIE 1931 2° color matching functions under D65** (visible 400–700 nm), then **bicubic downsampling** by a fixed scale factor **s**.  
- Dimensions: **h × w × 3**, where **h = H/s**, **w = W/s**.  
- The scale factor **s** (e.g., 2× or 4×) is provided in metadata and is consistent within a split.  
- Values are normalized to **[0,1]**.  
- Note: The LR RGB input contains no direct NIR (700–1000 nm) information; models must infer it from learned priors.

### 2. Hyperspectral Ground Truth (HSI-61)
- Ground truth hyperspectral cubes have **61 spectral bands**, covering **400–1000 nm** with **10 nm interval**.  
- These are derived by **subsampling from the original 448-band HSI cube**.  
- **Radiometric calibration** is applied to correct for sensor response and illumination, ensuring physically meaningful reflectance values.  

- Scenes are reflectance-normalized and captured under consistent indoor lighting near D65.

## Summary

- **Task:** Learn a mapping from **LR RGB (h×w×3) → HR HSI Cube (H×W×61)**, jointly upsampling spatially by **s** and expanding spectrally from **3 → 61** bands.  
- **Training set size:** 167 samples.  
- **Testing:** Combination of public, private, and hidden test sets ensures fair benchmarking; the same scale factor **s** is used for all test splits.