GAPA-3DGS: Dual-Branch Gaussian-Adaptive Perceptual Assessment for 3D Gaussian Splatting

Three-dimensional Gaussian Splatting (3DGS) has recently emerged as an efficient and photorealistic scene representation, offering real-time rendering with explicit control over geometry and appearance. However, perceptual quality assessment of 3DGS remains largely unexplored, as existing 2D or point-based metrics fail to capture the geometric coherence, attribute consistency, and cross-view perceptual fidelity of Gaussian fields. In this work, we present GAPA-3DGS, a Gaussian-Adaptive Perceptual Assessment framework specifically designed for no-reference 3DGS quality evaluation. Our method first integrates a geometry serialization-based preprocessing that converts unordered Gaussian primitives into structured sequences. It then employs a dual-branch encoder, comprising a Gaussian-adaptive point transformer for structural modeling and a Gaussian attribute-aware module for perceptual attribute learning, to jointly capture geometric and radiometric cues. An adaptive attribute-context fusion mechanism dynamically balances spatial and appearance information for perceptually aligned quality prediction. Extensive experiments demonstrate that GAPA-3DGS achieves superior correlation with human perception and outperforms conventional metrics across multiple 3DGS degradation scenarios.