Eve3D: Elevating Vision Models for Enhanced 3D Surface Reconstruction via Gaussian Splatting

We present Eve3D, a novel framework for dense 3D reconstruction based on 3DGaussian Splatting (3DGS). While most existing methods rely on imperfect priorsderived from pre-trained vision models, Eve3D fully leverages these priors byjointly optimizing both them and the 3DGS backbone.

This joint optimizationcreates a mutually reinforcing cycle: the priors enhance the quality of 3DGS, whichin turn refines the priors, further improving the reconstruction.

Additionally, Eve3Dintroduces a novel optimization step based on bundle adjustment, overcoming thelimitations of the highly local supervision in standard 3DGS pipelines.

Eve3Dachieves state-of-the-art results in surface reconstruction and novel view synthesison the Tanks & Temples, DTU, and Mip-NeRF360 datasets. while retaining fastconvergence, highlighting an unprecedented trade-off between accuracy and speed.