REIS: A Visual Analytics Tool for Rendering and Exploring Instance Segmentation of Point Clouds

3D Instance Segmentation (3DIS) of Point Clouds (PCs) is valuable for applications like autonomous vehicles, robotics, and Building Information Modeling (BIM). Current work on this topic is guided mainly by global metrics like mAP, which arguably do not support a deep, informed analysis of technique tradeoffs and, more importantly, directions for improvement. Qualitative analysis is widely adopted to provide such guidance, but it is generally implemented ad-hoc. This is true across many tasks in Deep Learning, but PC 3DIS is especially challenging to visually analyze due to the many variables involved: three spatial dimensions, colors, semantic labels, and instance IDs. We propose REIS, a visual analytics tool for Rendering and Exploring Instance Segmentation results. It supports qualitative analysis in two ways: first, through PC renderings targeted at efficient investigation of 3DIS results; second, by providing a systematic way to explore these results via the interactive Instance Detection Matrix- a confusion matrix analog that summarizes error and success cases, and allows the user to navigate through them. To show the efficacy of REIS, we use it to evaluate a state-of-the-art 3DIS approach on the S3DIS dataset. Our code is available at https://github.com/pedro-sidra/pcloud_explorer.