XSPLAIN: XAI-enabling Splat-based Prototype Learning for Attribute-aware INterpretability

3D Gaussian Splatting (3DGS) has rapidly become a standard for high-fidelity 3D reconstruction, yet its adoption in multiple critical domains is hindered by the lack of interpretability of the generation models as well as classification of the Splats. While explainability methods exist for other 3D representations, like point clouds, they typically rely on ambiguous saliency maps that fail to capture the volumetric coherence of Gaussian primitives. We introduce XSPLAIN, the first ante-hoc, prototype-based interpretability framework designed specifically for 3DGS classification. Our approach leverages a voxel-aggregated PointNet backbone and a novel, invertible orthogonal transformation that disentangles feature channels for interpretability while strictly preserving the original decision boundaries. Explanations are grounded in representative training examples, enabling intuitive ``this looks like that’’ reasoning without any degradation in classification performance. A rigorous user study (N=51) demonstrates a decisive preference for our approach: participants selected XSPLAIN explanations 48.4% of the time as the best, significantly outperforming baselines $(p<0.001)$, showing that XSPLAIN provides transparency and user trust. The source code for this work is available at: https://github.com/Solvro/ml-splat-xai

💡 Research Summary

XSPLAIN introduces the first ante‑hoc, prototype‑based explainable AI framework tailored for 3D Gaussian Splatting (3DGS) classification. The authors begin by noting that while 3DGS has become a dominant representation for high‑fidelity reconstruction, existing XAI methods for 3D data (e.g., point‑cloud saliency maps) fail to capture the volumetric coherence of Gaussian primitives, limiting their usefulness in safety‑critical domains.

The proposed system consists of three tightly coupled components. First, a PointNet‑inspired backbone processes the set of Gaussian primitives. Instead of the classic global max‑pooling, the backbone incorporates a voxel aggregation layer that partitions the normalized 3‑D space into a regular G³ grid. Each primitive is assigned a fixed voxel index based on its initial coordinates, and per‑voxel features are obtained by max‑pooling the point‑wise features within that voxel. This design preserves coarse spatial structure while retaining permutation invariance, enabling explanations that are anchored to concrete regions of the object.

Second, after the backbone has been trained for pure classification, its parameters are frozen. A learnable, invertible linear transformation matrix U∈ℝ^{C×C} is inserted between the voxel aggregation and the global pooling stage. U is constrained to be orthogonal, guaranteeing that the transformation does not alter the classifier’s decision boundaries. The orthogonal transformation disentangles the C latent channels so that each channel can be interpreted independently.

Third, a prototype‑based interpretability module leverages the disentangled channels. For each channel, the most representative training examples (prototypes) are identified based on similarity of the channel activations. During inference, the most active channels for a test sample are located, the corresponding voxels are highlighted, and the nearest prototypes are retrieved. Explanations are thus expressed as “this region looks like that region in prototype X,” providing intuitive, example‑driven reasoning grounded in both geometry and semantics.

Training proceeds in two stages. Stage 1 jointly optimizes the backbone with a standard cross‑entropy loss and a density‑aware regularization term. The regularizer aligns voxel activation magnitudes with the actual density of Gaussian primitives via a KL‑divergence between an activation distribution (softmax over ℓ₂ norms of voxel features) and a target density distribution derived from primitive counts. This encourages the model to focus on densely populated, geometrically meaningful voxels rather than sparse outliers. Stage 2 freezes the backbone and optimizes only U, using a purity objective that promotes channel‑wise prototype consistency while preserving orthogonality.

Empirical evaluation is performed on two 3DGS benchmarks: Shape‑Splat and MVImageNet‑GS. XSPLAIN matches or slightly exceeds the classification accuracy of baseline PointNet‑based 3DGS classifiers, demonstrating that the interpretability module does not sacrifice performance. Qualitative visualizations show that the highlighted voxels correspond to semantically meaningful object parts (e.g., chair legs, car wheels), and the retrieved prototypes are visually similar to those parts. Quantitative metrics—channel purity, prototype stability, and explanation consistency—are all superior to post‑hoc baselines such as LIME‑3D, Grad‑CAM, and random baselines.

A user study with 51 participants further validates the approach. Participants were shown explanations from XSPLAIN, LIME‑3D, Grad‑CAM, and a random baseline for the same inputs and asked to select the most understandable explanation. XSPLAIN was chosen 48.4 % of the time, a statistically significant advantage (p < 0.001). The study indicates that prototype‑based, example‑driven explanations are more intuitive and trustworthy for end‑users.

The paper’s contributions are threefold: (1) the first ante‑hoc, prototype‑based XAI method for 3DGS classification, (2) a novel combination of voxel aggregation and invertible orthogonal transformation that yields spatially coherent, semantically isolated explanations without degrading accuracy, and (3) a comprehensive evaluation—including quantitative benchmarks and a rigorous user study—demonstrating improved interpretability and maintained performance. Limitations include the exclusion of view‑dependent color information, sensitivity to voxel resolution and prototype count, and potential scalability challenges of orthogonal matrix learning in very high‑dimensional feature spaces. Future work may extend the framework to multimodal 3DGS inputs and explore adaptive voxelization strategies.