Factral/NeSpoF-segmentation

Extended NeSpoF Dataset

Fabian Perez¹² · Sara Rojas² · Carlos Hinojosa² · Hoover Rueda-Chacón¹ · Bernard Ghanem²

¹Universidad Industrial de Santander · ²King Abdullah University of Science and Technology (KAUST)

Introduction

This dataset is an extension of the NeSpoF dataset, enriched with ground-truth material labels for evaluating material segmentation in synthetic multi-view settings. The annotations provide consistent material labeling across different viewpoints for comprehensive scene analysis.

It is used in conjunction with UnMix-NeRF, a framework presented in the paper UnMix-NeRF: Spectral Unmixing Meets Neural Radiance Fields. UnMix-NeRF integrates spectral unmixing into Neural Radiance Fields (NeRF), enabling joint hyperspectral novel view synthesis and unsupervised material segmentation.

Dataset Sources

• Github: Official Code
• Paper: UnMix-NeRF: Spectral Unmixing Meets Neural Radiance Fields (ICCV 2025)
• Project Page: UnMix-NeRF Project Page
• Repository: Original NeSpoF Repository

Direct Use

This dataset is intended for training and evaluating models for material segmentation tasks, particularly useful for multi-view segmentation scenarios and NeRF-based material analysis.

Dataset Structure

The dataset has the following directory structure:

scene/
 ├── color/
 │    ├── eval/
 │    └── train/
 │         └── r_x.png
 └── raw/
      ├── eval/
      └── train/
           └── r_x.png

Here, x corresponds to the matching frame ID from the original NeSpoF dataset.

Dataset Creation

Source Data

Who are the source data producers?

The dataset extension was produced by the authors of the paper "UnMix-NeRF: Spectral Unmixing Meets Neural Radiance Fields," accepted at ICCV 2025.

Annotations

Annotation process

Annotations were automatically generated by rendering the ground-truth material indices, corresponding consistently across views and matching original scene frames.

Who are the annotators?

Automated rendering processed by mitsuba 3.

Bias, Risks, and Limitations

No known biases or risks are identified in this synthetic dataset. However, its synthetic nature may limit direct applicability to real-world scenarios without additional adaptation or fine-tuning.

Recommendations

Users should be aware that performance on this synthetic dataset may not fully generalize to real-world data without further adaptation.

Citation

If you use this dataset, please cite the following paper:

@inproceedings{perez2025unmix,
              title={UnMix-NeRF: Spectral Unmixing Meets Neural Radiance Fields},
              author={Perez, Fabian and Rojas, Sara and Hinojosa, Carlos and Rueda-Chac{\'o}n, Hoover and Ghanem, Bernard},
              booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
              year={2025}
            }

Dataset Card Contact

For inquiries regarding the dataset, please contact the corresponding authors listed in the referenced paper.