PADE-GS: a pixel-aware dynamic expansion for optimized 3D Gaussian splatting

Viewpoint optimization is essential for novel view synthesis (NVS), as it directly impacts rendering quality and visual performance. The 3D Gaussian point rendering (3DGS) has become a prominent technique in this domain due to its efficient rendering capabilities. However, existing methods often encounter difficulties in regions with sparse viewpoints, resulting in insufficient Gaussian point density and blurry distant views. To address these challenges, we introduce a pixel-aware dynamic expansion strategy for optimized 3D Gaussian splatting (PADE-GS). PADE-GS integrates a pixel-aware dynamic expansion technique (PADE-split) for adaptive point densification with a gradient scaling field to mitigate floaters. PADE-split enhances the point density in both sparse areas within and beyond the scene, utilizing pixel-aware viewpoint gradient optimization to achieve a more uniform distribution. To alleviate the occurrence of floating artifacts near the camera, caused by the over-generation of Gaussian points, we incorporate a depth-based gradient scaling field, significantly improving the visual quality. Experimental results demonstrate the effectiveness of our method across multiple datasets, showing clearer distant views and a substantial reduction in floaters in near-camera scenes.