SoftGroup for 3D Instance Segmentation on Point Clouds

SoftGroup++: Scalable 3D Instance Segmentation with Octree Pyramid Grouping

“…Recent state-of-the-art methods have made great strides in improving the performance of the segmentation predictions [38,39,44,21,26,13,20,37]. However, these methods still show limited scalability.…”

“…The results reveal that k-NN is the computational bottleneck, leading to the quick growth of inference time w.r.t. input size in HAIS [5] and SoftGroup [37]. This is because 3D-BoNet [44] HAIS [5] SoftGroup [37] SoftGroup++ (a) Runtime vs. number of points HAIS [5] SoftGroup [37] SoftGroup++ F1 score 3D-BoNet [44] HAIS [5] SoftGroup [37] SoftGroup++ (c) F1 score vs. runtime We measure the component time of processing a large scene of ~4.5M points.…”

“…input size in HAIS [5] and SoftGroup [37]. This is because 3D-BoNet [44] HAIS [5] SoftGroup [37] SoftGroup++ (a) Runtime vs. number of points HAIS [5] SoftGroup [37] SoftGroup++ F1 score 3D-BoNet [44] HAIS [5] SoftGroup [37] SoftGroup++ (c) F1 score vs. runtime We measure the component time of processing a large scene of ~4.5M points. The measurement exposes k-NN as computational bottleneck of existing methods.…”

“…SSTNet [20] constructs a semantic superpoint tree that is traversed and split at intermediate nodes to obtain object instances. SoftGroup [37] proposes to perform grouping on soft semantic scores to address the problem of the hard semantic predictions that propagates the errors to instance predictions. Despite improvements in accurateness, these methods are generally slow to handle large scenes.…”

“…Then, the segmentation results from different blocks are merged using the BlockMerging algorithm [38]. In [13,20,37], large scenes are divided into 4 parts, and then the predictions are merged to obtain the final results. The divideand-merge strategy exhibits various drawbacks.…”