Surface Reconstruction From Point Clouds via Image-Free Point-to-Gaussian Inference.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• Surface reconstruction from point clouds traditionally prioritizes geometric accuracy over texture.
• Existing methods often fail to produce textured meshes when only 3D point data is available.

Purpose of the Study:

• To develop a novel method for high-quality, colored surface reconstruction from point clouds without using images.
• To introduce a new perspective by representing points as Gaussian splats for enhanced reconstruction.

Main Methods:

• A universal Point-to-Gaussian model is trained to infer Gaussian splat attributes from point cloud data.
• Free-form Gaussian rendering is utilized for obtaining colored surfaces.
• Truncated Signed Distance Function (TSDF) fusion is employed with virtually rendered images and depth maps.

Main Results:

• The proposed method successfully recovers object appearance and adheres closely to the input shape.
• High-quality, colored meshes are generated, surpassing state-of-the-art methods in reconstruction metrics.
• The approach demonstrates efficiency and simplicity in the reconstruction process.

Conclusions:

• This novel Gaussian splat-based approach offers a significant advancement in point cloud surface reconstruction.
• The method effectively generates textured and geometrically accurate 3D meshes from point clouds.
• It provides a robust and efficient solution for creating detailed 3D models.