README.md

---
license: apache-2.0
language:
- en
size_categories:
- 1K<n<10K
task_categories:
- image-segmentation
tags:
- 3d-reconstruction
- artifact-detection
- image-quality-assessment
- human-annotation
---

# <img src="https://www.svgrepo.com/show/510149/puzzle-piece.svg" width="22"/> Puzzle Similarity

[Project page](https://nihermann.github.io/puzzlesim/) | [Paper](https://arxiv.org/abs/2411.17489) | [Code](https://github.com/nihermann/PuzzleSim)

-----

> This repository contains the dataset presented in the ICCV 2025 paper "Puzzle Similarity: A Perceptually-guided Cross-Reference Metric for Artifact Detection in 3D Scene Reconstructions"  
> Authors: Nicolai Hermann, Jorge Condor, and Piotr Didyk

### Dataset Description
The Dataset consists of 36 hand-selected 3D Gaussian Splatting renderings containing common reconstruction artefacts, ground truths, human-annotated masks, and a set of reference views of the same scene.

Each mask is an average of 22 binary masks, each created by a different human participant who was asked to annotate areas in the reconstructed images that they perceived as visually degraded, unnatural, or incongruent. The dataset can be used to benchmark No-Reference, Cross-Reference, and Full-Reference image quality metrics for their correlation with human judgment. The naming convention of the data is as follows:

- `dataset_perc_id_mask.png` (grayscale)
- `dataset_perc_id_artifact.png`
- `dataset_perc_id_gt.png`
- `dataset_perc_refs/`

The dataset was created by fitting 3DGS to a scene while using a reduced number of training views. We withheld a percentage of views (perc) and added them to the validation dataset, which is found in the *_refs/ directory for each respective sample to act as unseen reference views for Cross-Reference metrics. We fitted the scenes while withholding 60%, 70%, or 80% to get a wider variety and strength of artifacts. (Disclaimer: perc actually refers to proportions, so the possible values are 0.6, 0.7, or 0.8)

The included datasets are a collection from the Mip-NeRF360 [1], Tanks and Temples [2], and Deep Blending [3] datasets; thus, the ground truths are copies from their data.

[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter 594 Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. 595 Mip-NeRF: A Multiscale Representation for Anti-Aliasing 596 Neural Radiance Fields, 2021.

[2] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen 656 Koltun. Tanks and temples: benchmarking large-scale scene 657 reconstruction. ACM Transactions on Graphics, 36(4):1–13, 658 2017

[3] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, 619 George Drettakis, and Gabriel Brostow. Deep blending for 620 free-viewpoint image-based rendering. ACM Transactions 621 on Graphics, 37(6):1–15, 2018.

### Citation
If you find this work useful, please consider citing:
```bibtex
@inproceedings{hermann2025puzzlesim,
      title={Puzzle Similarity: A Perceptually-Guided Cross-Reference Metric for Artifact Detection in 3D Scene Reconstructions},
      author={Nicolai Hermann and Jorge Condor and Piotr Didyk},
      booktitle={ICCV},
      year={2025},
}
```