Segmenting candidate-interaction regions: Preprocessing 3D point clouds for interactive tabletop implementations

Easy access to depth sensors has promoted exploring how point clouds can be leveraged to augment tabletops in the everyday context. However, point-cloud operations are computationally expensive and challenging to perform in real-time, particularly absent dedicated GPU-compute potential. In this paper, we propose mitigating the high computational costs by segmenting candidate-interaction regions near real-time. Focusing on CPU-based architectures, we put forward a modular pipeline that minimizes point-cloud volumes and proposes possible regions of interest for interactive tabletop implementations. While emphasizing a general solution adaptable for variable depth cameras, we also center on modular design for flexible pipeline filters. Our validation is two-fold. First, we employ a popular depth camera and apply the pipeline over a unique scene to establish initial performance measures. Second, we employ the XXXX dataset and report the pipeline’s performance over variable scenes from the dataset. For the unique scene, our initial findings indicate that point-cloud data volumes are reduced by up to 70\%, segmenting candidate-interaction regions under 35 ms. For the XXXX database, data volumes are reduced by an average of 70\%, and on average, candidate-interaction regions are segmented under 35 ms. Going one step further, we condense the approach into an open-source solution and conclude the paper by elaborating on the benefits of the segmenting candidate-interaction regions near real-time.