Conceptual

EasySplat: View-Adaptive Initialization and KNN Densification for 3D Gaussian Splatting

EasySplat is a 3D Gaussian Splatting framework for novel view synthesis that replaces noisy SfM/COLMAP initialization with pointmap priors from a Transformer MVS model (DUSt3R). To keep dense-view pointmap construction tractable it groups the image sequence by adjacent-frame cosine similarity, estimating only intra-group pairwise pointmaps before global alignment into camera poses and a point cloud. Its KNN-based densification finds each Gaussian's nearest neighbors, computes their mean covariance, and splits any Gaussian exceeding that neighborhood shape, growing detail where the initial cloud is sparse. It reports SOTA rendering and pose accuracy on Tanks&Temples and CO3DV2.