ActiveTwig: Morphology-Guided Active Reconstruction for Slender Structures Using Gaussian Splatting

Yuchuan Jin, Cunjia Liu

Keywords: Precision Agriculture, Agricultural Automation, Active Perception, Next-Best-View Planning, Gaussian Splatting

TL;DR: An anisotropy-aware active perception method that improves reconstruction of thin, occluded branches by prioritizing informative viewpoints.

Abstract: Accurate 3D reconstruction of slender branches is critical for robotic pruning, yet thin and occluded geometries challenge existing active perception systems. Current Gaussian Splatting (GS)-based methods typically rely on isotropic confidence metrics, leading to inefficient viewpoint allocation for tubular structures. We propose ActiveTwig, a morphology-guided active reconstruction framework that incorporates geometric anisotropy into viewpoint evaluation. By extracting the principal axis from GS primitives, we introduce a direction-aware confidence model that prioritizes circumferential observations. This is integrated into a target-centric Next-Best-View (NBV) strategy with ROI-constrained sampling and depth-aware weighting. Simulation results on TreeNet3D show that ActiveTwig improves skeleton accuracy by 17–27% and achieves up to 100% success rate in topology completion, outperforming GS-based and coverage-driven baselines.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 5