Targeted Structure Completion for Sparse-View 3D Reconstruction in Autonomous Driving

FocusGS targets structural incompleteness in sparse-view 3D reconstruction for autonomous driving. Instead of uniformly densifying the entire volume, it derives a 3D Geometric Ambiguity Manifold to identify regions affected by occlusion or high geometric uncertainty. A lightweight targeted structure completion module then instantiates and optimizes continuous Gaussian queries only inside these ambiguous sparse topological regions. According to the abstract, FocusGS improves the efficiency-quality trade-off on driving-centric benchmarks, reduces the total number of Gaussians by approximately 74%, and decreases rendering time by approximately 34%.