PI-Splatting: Perceptual Importance-guided Compression for Efficient Gaussian Splatting

Abstract

3D Gaussian Splatting (3DGS) has recently emerged as a prominent paradigm for novel view synthesis due to its high-fidelity and real-time rendering capabilities. However, its explicit scene representation typically requires a large number of Gaussian primitives, leading to substantial storage overhead and significant redundancy. Existing pruning-based compression methods typically use heuristic metrics or learnable masks to identify primitives for removal, but it remains challenging to achieve both aggressive and adaptive pruning. To address this issue, we propose PI-Splatting, a perceptual-importance-guided adaptive-pruning framework for compact 3DGS representations. Specifically, we introduce perceptual importance, a novel metric for characterizing the reconstruction significance of spatial regions, derived from visibility and visual saliency cues inspired by human perception. Based on this metric, we propose perceptual importance-guided stochastic masking that assigns different masking probabilities to Gaussian primitives during training, preserving the representational capacity of perceptually important regions while suppressing less important ones. In the later stage of optimization, primitives with low perceptual importance are pruned, resulting in a more compact representation with minimal degradation in reconstruction quality. Extensive experiments demonstrate that PI-Splatting significantly reduces the number of Gaussian primitives while preserving rendering quality, and outperforms state-of-the-art methods in both pruning aggressiveness and adaptivity.