MultiScan Benchmarks

We systematically carried out following 3 tasks on MultiScan dataset: 3D instance segmentation of objects and parts, and articulated object mobility prediction. We evaluated recent methods for each task.

Object Instance Segmentation

We benchmark point-cloud based methods Jiang et al. [jiang2020pointgroup], Liang et al. [liang2021instance], Chen et al. [chen2021hierarchical] on our MultiScan dataset. We select the top 17 most common object categories, excluding room architecture (floor, wall, beam, etc.) and small objects.

Part Instance Segmentation

We benchmark the same 3 point-cloud based methods on our MultiScan dataset for part instance segmentation. We select the top 5 most common part categories (static, door, drawer, window, lid).

Mobility Estimation

We compare two category-agnostic methods that operate on point clouds: Shape2Motion (S2M) Wang et al. [wang2019shape2motion] and OPDPN Jiang et al. [jiang2022opd]. Both use a PointNet++ backbone. Given an input point cloud, we predict the part for each point, and the motion parameters for each part, by outputting the set of moving parts and their joint parameters. We assume each moving part is only associated with one joint.