RWLQ Dataset: Benchmarking 3D Artifact Suppression

• RWLQ dataset is a benchmark for evaluating artifact suppression in 3D Gaussian Splatting using real-world, low-quality imagery.
• It captures challenges such as missing viewpoints, low resolution, and noise, enabling robust testing of 3D reconstruction methods.
• Techniques like EFA-GS show improved PSNR and natural geometry, highlighting the dataset’s utility in advancing reconstruction algorithms.

The RWLQ dataset refers to a benchmark introduced for evaluating artifact suppression in 3D Gaussian Splatting (3DGS) under real-world low-quality image and initialization conditions. It is principally associated with the work on Eliminating-Floating-Artifacts Gaussian Splatting (EFA-GS), which directly addresses floating artifacts in 3D reconstructions learned from incomplete, noisy, or suboptimal cellphone imagery. The dataset provides a challenging environment for scene-level 3D neural representations due to missing viewpoints, low resolution, and the prevalence of noise—factors that expose the limitations of existing 3DGS techniques and motivate algorithmic improvements.

1. Dataset Definition and Motivations

The RWLQ dataset comprises real-world scenes captured with consumer-grade or mobile device cameras, resulting in images that lack comprehensive coverage as well as high-resolution detail. The dataset is specifically engineered to contain the characteristic problems of real-world data encountered in practical 3D capture scenarios: missing viewpoints, low per-frame quality, and significant noise or uncertainty in initial estimates of scene geometry and appearance. These deficiencies are known to induce "floating artifacts"—spurious, detached geometry during 3D Gaussian optimization—that degrade visual fidelity in classical 3DGS pipelines.

The motivation for establishing the RWLQ dataset was to facilitate rigorous, quantitative, and qualitative analyses of artifact formation and suppression methods in 3DGS under realistic constraints, with explicit tracking of PSNR, SSIM, and perceptual metrics, and with a qualitative focus on artifact prevalence in renderings.

2. Data Collection and Scene Structure

Scenes in the RWLQ dataset are constructed from image collections obtained under non-laboratory, uncontrolled conditions. The source imagery is captured with widely-available devices such as cellphone cameras, introducing lower effective resolution and increased levels of noise. Crucially, ground truth geometry and camera poses may be partially missing or only available through estimation processes, further emulating practical reconstruction settings.

Within each scene, the following components are typically present:

• RGB images exhibiting low resolution and compression artifacts;
• Incomplete viewpoint coverage, resulting in occluded or unobserved geometric regions;
• Initialization noise for both spatial positions and appearance attributes of the 3D Gaussians.

No explicit quantitative breakdown of scene count, number of images per scene, or calibration statistics is provided, but the defining properties are the deliberate inclusion of realistic low-quality factors.

3. Benchmarking Artifact Suppression: Experimental Protocols

The dataset’s central function is to benchmark artifact suppression algorithms in 3DGS, with a particular focus on methods like EFA-GS that operate in the frequency domain to manage spatial-frequency content during scene optimization. The required protocol comprises the following:

• Training 3D Gaussian Splatting representations on the provided images, using noisy, real-world-scale initializations;
• Evaluating reconstructions for both visual fidelity and reduction in floating artifact prevalence;
• Quantifying improvements using standard metrics such as PSNR (Peak Signal-to-Noise Ratio), SSIM, and LPIPS, with a focus on the delta versus vanilla 3DGS or Mip-splatting baselines.

Experiments on the RWLQ dataset reported an improvement of approximately 1.68 dB in PSNR with the EFA-GS(Mip) method relative to Mip-splatting. Qualitative analysis confirmed more natural reconstructions, with fewer detached structures in the scene geometry.

4. Algorithmic Implications: Frequency-Domain Artifact Analysis

Methods evaluated on the RWLQ dataset, such as EFA-GS, are compelled by its properties to address the frequency-domain mismatch that arises during noisy or incomplete initialization. The underlying insight is that under-optimized or “over-shrunk” Gaussians contribute disproportionately to high-frequency content, failing to robustly encode low-frequency scene structures. This effect is formalized via the Fourier transform of individual Gaussians:

$$|\mathcal{F}(G(x), \omega)| = \exp\left(-\frac{1}{2} \omega^\top \Sigma\, \omega\right)$$

where $\Sigma$ is the covariance matrix of the Gaussian. In RWLQ conditions, Gaussian expansion strategies, depth-based interpolation, and anisotropic scaling mechanisms are applied dynamically based on per-Gaussian gradient signals and view-based depth metrics to ensure sufficient low-frequency encoding before iteratively recovering high-frequency detail.

5. Applications in 3D Editing and Practical Reconstruction

Beyond artifact suppression, success on the RWLQ dataset determines algorithmic suitability for downstream and user-facing applications—including 3D scene editing and real-time manipulation. Methods robust to the RWLQ challenges produce cleaner geometry and fewer refractive artifacts, enabling easier attribute editing in pipelines such as GaussianEditor. The ability to execute semantic or attribute-changing commands on reconstructed geometry without incurring artifact propagation or geometry corruption has been experimentally demonstrated. Artifact suppression is essential to preserve both object boundaries and semantic consistency during scene edits.

6. Significance and Broader Impact

The RWLQ dataset fills a critical gap by providing a realistic, challenge-focused benchmark for evaluating and advancing neural 3D representations in non-idealized conditions. Its emphasis on challenging input regimes exposes failure modes not captured by academic, high-quality datasets. Improvements evidenced on RWLQ directly translate to increased robustness and perceptual quality in consumer, industrial, and creative applications that rely on monocular or limited multi-view image capture. The frequency-domain perspective and associated suppression strategies catalyzed by the RWLQ benchmark are poised for further generalization in broader 3D vision research.