Compact Feed-Forward 3D Gaussians via Saliency-Guided Primitive Merging
The method reduces redundant 3D Gaussian primitives from feed-forward methods by implementing an encode-merge-decode pipeline. It uses saliency-guided superpixel segmentation to group similar Gaussians, a learned encoder to create compact Feature Gaussians, and a cross-view matching module to merge redundant representations. This results in a structure that retains high visual quality while drastically cutting the number of primitives.