UniqueSplat: View-Conditioned 3D Gaussian Splatting for Generalizable 3D Reconstruction

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Existing feed-forward 3D Gaussian Splatting methods generate fixed Gaussians, limiting viewpoint adaptability.
• These methods do not incorporate target view information, hindering customized 3D radiance field reconstruction.

Purpose of the Study:

• To propose UniqueSplat, a view-conditioned feed-forward 3D Gaussian Splatting model.
• To enable dynamic adjustment of Gaussians based on view-specific information for improved 3D scene reconstruction.

Main Methods:

• Developed a view-conditioned feed-forward 3D Gaussian Splatting model named UniqueSplat.
• Introduced a two-branch view-conditioned hyperNetwork to learn both view-agnostic embeddings and view-specific knowledge.
• Incorporated view-conditioned information as a prior into network parameters for dynamic Gaussian adjustment.

Main Results:

• UniqueSplat demonstrates superior performance over state-of-the-art methods on RealEstate10K, ACID, and DTU datasets.
• Achieved state-of-the-art results in cross-dataset evaluation, highlighting strong generalization capabilities.
• Successfully reconstructs customized 3D radiance fields tailored to specific view queries.

Conclusions:

• UniqueSplat effectively addresses the limitations of fixed- सेGaussian approaches in 3D Gaussian Splatting.
• The proposed view-conditioned hyperNetwork enables dynamic adaptation to specific views, enhancing reconstruction quality.
• The model exhibits excellent generalization ability, outperforming existing methods in cross-dataset scenarios.