Research map / 3D vision / NeRF and radiance fields
Novel Class Discovery for Representation of Real-World Heritage Data as Neural Radiance Fields (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NCD-NeRF, a two-stage pipeline that uses novel class discovery to cluster unlabeled multi-site heritage images and then trains NeRF representations per discovered class, evaluated on a custom UNESCO heritage dataset from India.
Problem
Representing unstructured, unlabeled real-world heritage image collections containing multiple sites as NeRFs requires costly manual labeling and organization of classes.
Approach
Combine novel class discovery (SMILE-style clustering of unlabeled images using a labeled-set pretrained model) with per-class NeRF reconstruction, so that multiple unlabeled heritage sites are automatically separated and rendered in parallel without human labeling.
Key results
- SMILE achieves 64.32% All accuracy on IDH novel-class discovery, the highest among DTC (30.7), Restune (45.31), Class-INCD (55.89), and OCD (64.32) in Table 1.
- NeRF renderings for four heritage sites reach PSNR around 21.8 to 22.8 and SSIM 0.58 to 0.69 (Table 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SMILE for novel-class discovery instead of other evaluated NCD methods | NCD-NeRF · IDH heritage data | Strict-Hungarian accuracy | Helped |
| Use novel class discovery instead of trivial unsupervised alternatives such as K-Means | NCD-NeRF · IDH heritage data | quantitative performance | Helped |
Limitations
- Evaluation limited to a single custom dataset of four to nine heritage sites
- No ablation or comparison of NeRF quality against unsupervised clustering alternatives
- Table 2 reports NeRF metrics without baseline comparisons
Experimental setup
Datasets: IDH (custom UNESCO World Heritage India dataset)
Benchmarks: IDH
Metrics: Strict-Hungarian accuracy (Old/New/All) SSIM PSNR LPIPS
Baselines: DTC Restune Class-INCD OCD K-Means
Builds on: NeRF SMILE Class-INCD OCD
Closest papers by meaning
- NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections (CVPR 2021 · similarity 0.89)
- D-NeRF: Neural Radiance Fields for Dynamic Scenes (CVPR 2021 · similarity 0.88)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.86)
- Ced-NeRF: A Compact and Efficient Method for Dynamic Neural Radiance Fields (AAAI 2024 · similarity 0.86)
- Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates (CVPR 2024 · similarity 0.86)
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.