This thesis proposes a novel active 3D object reconstruction framework under ground environments designed to handle the challenges of real-world environments with frequent and diverse occlusions. Our approach employs a viewpoint selection policy that ...
This thesis proposes a novel active 3D object reconstruction framework under ground environments designed to handle the challenges of real-world environments with frequent and diverse occlusions. Our approach employs a viewpoint selection policy that leverages amodal segmentation features, allowing the robot to reason about partially occluded objects and select informative viewpoints that improve reconstruction quality while minimizing trajectory length. To guide the robot toward observing occluded regions, we introduce an occlusion-aware reward using amodal features that encourages uncovering hidden surfaces. For training and evaluation, we construct two occlusion-centric benchmark datasets (Occ-House3K and Occ-OmniObject3D) within the simulated environment. Our method achieves superior reconstruction coverage and path efficiency compared to existing state-of-the-art methods originally designed for aerial views, with coverage ratio improvements of approximately 2.04% and 3.05%, and trajectory length reductions of up to 45.04 meters. These results demonstrate a robust, embodied solution for scene understanding from a ground-view perspective.