Dynamic Scene Reconstruction and Understanding

Traditional approaches to 3D reconstruction have achieved remarkable progress in static scene acquisition. The acquired data serves as priors or benchmarks for many vision and graphics tasks, such as object detection and robotic navigation. Thus, obtaining interpretable and editable representations from a raw monocular RGB-D video sequence is an outstanding goal in scene understanding. However, acquiring an interpretable representation becomes significantly more challenging when a scene contains dynamic activities; for example, a moving camera, rigid object movement, and non-rigid motions. These dynamic scene elements introduce a scene factorization problem, i.e., dividing a scene into elements and jointly estimating elements’ motion and geometry. Moreover, the monocular setting brings in the problems of tracking and fusing partially occluded objects as they are scanned from one viewpoint at a time. This thesis explores several ideas for acquiring an interpretable model in dynamic environments. Firstly, we utilize synthetic assets such as floor plans and object meshes to generate dynamic data for training and evaluation. Then, we explore the idea of learning geometry priors with an instance segmentation module, which predicts the location and grouping of indoor objects. We use the learned geometry priors to infer the occluded object geometry for tracking and reconstruction. While instance segmentation modules usually have a generalization issue, i.e., struggling to handle unknown objects, we observed that the empty space information in the background geometry is more reliable for detecting moving objects. Thus, we proposed a segmentation-by-reconstruction strategy for acquiring rigidly-moving objects and backgrounds. Finally, we present a novel neural representation to learn a factorized scene representation, reconstructing every dynamic element. The proposed model supports both rigid and non-rigid motions without pre-trained templates. We demonstrate that our systems and representation improve the reconstruction quality on synthetic test sets and real-world scans