Center-Aware Instance Segmentation for Close Objects in Industrial 3-D Point Cloud Scenes

Segmenting small-scale and close 3-D objects from a complicated 3-D point cloud scene is a crucial yet challenging task for industrial scene understanding. Previous 3-D instance segmentation work (such as PointGroup and SoftGroup) suffers from limited accuracy when dealing with objects in close proximity. This article proposes a center-aware 3-D instance segmentation framework that performs two cooperative stages, including bottom-up center-guided instance proposal generation and top-down center-aware instance mask refinement, to jointly address the challenges of segmenting close objects. In the first stage, the object clusters are first segmented from the cluttered background. We propose to determine the center of each object by predicting its
k
-nearest neighbors with their respective offsets, which enables an instance proposal generated for each object. In the second stage, center-based relative position encoding is performed on all points in each instance proposal. We concatenate the explicitly encoded geometric features (low-level) with high-level semantic features and construct the center-aware feature representation for instance mask prediction. We conducted experiments on a real-world industrial point cloud dataset that involves the important aero-engine component—close clamps in the cluttered scenario. The results show that our proposed method achieves state-of-the-art overall 3-D instance segmentation precision with
$\mathbf{AP}_{\mathbf{avg}}$
of 0.905 and
$\mathbf{AP}_{\mathbf{90}}$
of 0.736.