Implicit Surface Reconstruction from Sparse and Noisy Poses with Large Motions

ByeongIn Joung, Youngho Yoon, Hyun-Kurl Jang, Kuk-Jin Yoon

Keywords: multi-view reconstruction, sparse input views, noisy poses, implicit surface reconstruction

TL;DR: By proposing uncertainty-aware consistency losses and implicit surface smoothing method, we successfully trained a model for neural implicit surface reconstruction with sparse and noisy input views.

Abstract: Recent advances in implicit surface reconstruction have significantly improved 3D reconstruction techniques. However, challenges persist, particularly when dealing with sparse and noisy poses. Traditional methods attempted to address these challenges through photometric and geometric consistency, but they often struggled as camera baselines increased. This difficulty arises due to incorrect guidance caused by occlusions during the learning of neural implicit representation. To overcome this issue, we propose an approach that incorporates uncertainty-aware guidance for multi-view consistency, allowing for better adaptation to scenarios with sparse and noisy inputs. Additionally, to facilitate the learning of surface geometry in a challenging setup, we propose a geometric smoothing termed progressive SDF loss. Through empirical studies on occlusion handling and geometric smoothing, our method achieved state-of-the-art performance, significantly enhancing both the refinement of noisy camera poses and surface reconstruction quality. This advancement strengthens the robustness and flexibility of implicit surface reconstruction in challenging conditions, paving the way for more effective applications in computer vision and 3D scene understanding.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 5407