Camera-Agnostic Pruning of 3D Gaussian Splats via Descriptor-Based Beta Evidence

The camera-agnostic formulation is principled and timely for MPEG I-3DGS interoperability. The HSFH descriptor extends FPFH naturally by incorporating spherical harmonics power spectra and histogram representations, providing a compact signature of local appearance variation. The Beta evidence framework credibly balances pruning likelihood ($\mu_i=B_i/(A_i+B_i)$) with uncertainty ($\sigma_i^2$), and the ablation study confirms that both descriptor-based modeling and Beta uncertainty estimation contribute meaningfully to quality retention. The evaluation on held-out camera views (following MPEG CTC protocol) represents fair assessment practice.

Several empirical choices raise reproducibility and methodological concerns. The evidence aggregation weights in Eq (4) (0.50, 0.35, 0.20, etc.) are stated to be "selected empirically" without systematic justification or sensitivity analysis. The "optimistic confidence" heuristic in Eq (7) with fixed $\gamma=0.25$ is arbitrarily chosen to "softly reward uncertainty," yet this formulation directly counteracts the conservative interpretation of uncertainty in Eq (6) without theoretical grounding. The comparison baselines are unfairly handicapped: LightGSPrune and ConfSplatPrune are truncated "pruning-only variants" isolated from full compression pipelines, removing recovery optimization and quantization that significantly impact rate-distortion trade-offs. This makes claims about outperforming camera-dependent methods at high pruning ratios misleading, as the full LightGaussian system achieves 15$\times$ compression with quality preservation, whereas this method only tests up to 30% pruning without considering bitrate. Additionally, the spherical harmonics coefficients used in the appearance component are themselves view-dependent representations, somewhat undermining the "camera-agnostic" framing despite the authors' claims of operating without camera parameters.

The quantitative comparison in Table 1 reveals that BetaDescPrune remains competitive but does not clearly dominate camera-aware methods. On bartender (tracked), camera-aware methods achieve PSNR 89.81-89.83 dB at low pruning versus 88.52 dB for the proposed method—a noticeable gap indicating that camera information provides non-negligible utility. Conversely, on object-centric plant sequences, all methods saturate near 97 dB PSNR, rendering the comparison uninformative. The claim that the method achieves "slightly better performance" at high pruning applies only to specific sequences (breakfast, cinema) and masks inconsistent performance elsewhere. Crucially, the paper measures only pruning ratio without reporting final bitrate or storage costs, making it impossible to assess compression efficiency relative to full pipelines like LightGaussian that achieve 15$\times$ size reduction.

Reproducibility is moderately impaired by missing implementation details and empirical design choices. No code or data release URLs are provided. The method depends on critical hyperparameters ($\gamma=0.25$, aggregation weights in Eq 4, voxel size 1-2% of bounding box diagonal) selected without systematic protocols. The voxelized downsampling step introduces interpolation complexity (distance-based weighting from voxels to splats) that is underspecified. Computational costs of HSFH descriptor extraction—particularly for spherical harmonics power spectrum computation and nearest-neighbor searches in large scenes—are unreported, making scalability assessment impossible. While the MPEG CTC dataset is standardized, access may be restricted, and without code release, independent reproduction of the exact Beta evidence formulation and HSFH encoding remains challenging.