A pr 2 01 9 Graph based Dynamic Segmentation of Generic Objects in 3 D ∗

3D segmentation is a promising building block for high level applications such as scene understanding and interaction analysis. New challenges emerge for computer vision techniques in generic scenarios with RGBD stream data. We focus on temporally evolving 3D point clouds in order to segment an image sequence into regions, which should ideally correspond to meaningful objects in the scene. To achieve this goal, some approaches incorporate high level knowledge into the segmentation process, such as object models [2] and accurate object annotations in the initialization stage [7]. However, most computer vision applications involve large amounts of data with different types of scenes containing several objects, which makes those methods difficult to be adapted to generic scenes. In order to segment generic objects in the 3D scene, without explicit object models or accurate initialization, we will consider that objects correspond to ”compact point clouds” in the 3D-space plus time domain. However, point clouds corresponding to an object can break into different connected components due to occlusions, or can merge with point clouds corresponding to other objects, producing a single connected component, when they become spatially close (object interaction). Our system produces a robust spatio-temporal segmentation of the point clouds, analyzing their connectivity to define the objects according to the evidence observed up to a given temporal point. To tackle spatio-temporal connectivity Hickson et al.[3] propose a graph based model to hierarchically perform a video segmentation from over-segmented frames. But the over-segmentation in different frames is calculated independently, which may lead to a temporal consistency problem. Abramov et al.[1] propose to transfer labels from frame to frame, relying on optical flow. But this requires a good initialization and is restricted to the performance of the opti-