BGLab/HSI-SC-NeRF

---
license: cc-by-nc-4.0
task_categories:
- image-to-3d
---

# HSI-SC-NeRF: NeRF-based Hyperspectral 3D Reconstruction using a Stationary Camera for Agricultural Applications

[**📄 Paper Link**](https://arxiv.org/abs/2602.16950)

## 🧾 Overview

This dataset accompanies our work on **HSI-SC-NeRF**, a stationary-camera-based hyperspectral NeRF framework for 3D plant phenotyping and postharvest agricultural inspection. Unlike conventional NeRF pipelines that require camera motion around a static object, our approach uses a **fixed camera and a rotating object**, making the acquisition setup simpler and more practical for controlled imaging. The system is built around a custom **PTFE (Teflon) studio chamber** designed to provide diffuse and spatially uniform illumination for hyperspectral acquisition. After white-reference-based spectral calibration and COLMAP-based pose estimation, a multi-channel NeRF jointly reconstructs all spectral bands to produce high-fidelity 3D hyperspectral point clouds that preserve both geometric structure and spectral reflectance information.

The dataset includes white-reference data for spectral calibration, raw hyperspectral acquisitions, an iPhone-captured video for spatial validation, pseudo-RGB images for pose estimation, COLMAP camera poses, masks for fine-tuning, two-stage NeRF training outputs, evaluation metric JSON files and per-band error maps for spectral validation, and final hyperspectral 3D point clouds. It is intended to support reproducible research in hyperspectral 3D reconstruction, spectral-geometric scene representation, and automated agricultural inspection under controlled imaging conditions.

## ✨ Key Contributions

- Stationary-camera NeRF for 3D hyperspectral reconstruction without requiring camera movement
- A custom PTFE studio chamber for diffuse and spatially uniform hyperspectral illumination
- A multi-channel NeRF formulation that jointly reconstructs all hyperspectral bands
- High-fidelity 3D hyperspectral point clouds through joint spectral-geometric reconstruction
- Validation for automated, high-throughput postharvest agricultural inspection

## 📁 Dataset Directory Structure

For readability, the directory tree below is shown in a simplified logical form. Actual released folder names may retain some original pipeline-specific naming for compatibility with preprocessing and training scripts.

```text
HSI-SC-NeRF/
├── wr/
├── raw/
└── pcd/
```

## wr/

Contains white reference (WR) data used for spectral calibration.

## raw/

Contains raw acquisition data for three objects:

- Apple (bruised)
- Maize ear
- Pear

For each object, the raw hyperspectral acquisition includes .dat, .hdr, and preview image files. In addition, the maize sample includes an iPhone video for spatial validation:

- Oh43x_SC: stationary camera + rotating object (target configuration)
- Oh43x_gt.MOV: moving camera + stationary object (reference video for spatial validation)

## pcd/

Contains the final exported 3D hyperspectral point clouds (204 bands; 1 million points per object), generated from the fine-tuned models with λ_ang = 0.25 and λ_hsi = 0.75.

## 🧠 HSI-SC-NeRF Pipeline

The reconstruction workflow consists of three main stages:

1. Dataset Acquisition
Experimental setup and multi-view hyperspectral image collection using a stationary camera and a rotating object.

2. Data Preprocessing
White-reference spectral calibration, pseudo-RGB generation, and COLMAP-based pose estimation.

3. NeRF-Based Hyperspectral Point Cloud Reconstruction
Multi-channel hyperspectral NeRF training, hyperspectral point cloud generation, and refinement.

This pipeline produces final 3D hyperspectral point clouds that support downstream spatial and spectral analysis.

## 🌱 Applications

- Hyperspectral 3D reconstruction
- Plant phenotyping
- Postharvest agricultural inspection
- Spectral-geometric representation learning
- Benchmarking stationary-camera NeRF pipelines under controlled imaging conditions

## 📜 License

**CC-BY-NC-4.0**
[Creative Commons Attribution-NonCommercial 4.0 International License](https://creativecommons.org/licenses/by-nc/4.0/)

## 🔖 Citation
If you use this dataset in your work, please cite:

```text
@article{ku2026hyperstationarynerf,\
title = {HSI-SC-NeRF: NeRF-based Hyperspectral 3D Reconstruction using a Stationary Camera for Agricultural Applications},\
author = {Kibon Ku, Talukder Z. Jubery, Adarsh Krishnamurthy, Baskar Ganapathysubramanian},\
year = {2026},\
journal = {arXiv preprint arXiv:2602.16950}\
}
```