Search

3D object detection is important for various indoor applications to understand the environment. Previous voting-based methods voted on the center of each seed point, which may suffer from errors from background points or adjacent objects. And the size-fixed feature grouping module is unsuitable for indoor objects with variable sizes. In this paper, we propose a Size Adaptive Clustering method for 3D object detection in point clouds . First, we present a super-voting module to divide seed points into foreground and background points and perform enhanced voting on the foreground seeds. To create a good match for the feature clustering area and the size of an object, we design a size-adaptive clustering module to infer a clustering radius based on the seed-to-vote displacement offset. Finally, because indoor objects are highly related to spatial room layouts, a position-aware module is used to calculate aware weights among objects and enhance the features of occluded objects. Experiments show that our method outperforms VoteNet by a large margin on ScanNet V2 (mAP@0.250 +8.3%, mAP@0.50 +14.2%) and SUN RGB-D (mAP@0.250 +3.5%, mAP@0.50 +13.6%). The proposed method can detect indoor objects with variable sizes in high accuracy, and perform robustly in case of occluded objects. The code of 3DSAC will be available at github-3DSAC.