Transformer-based 3D Instance Segmentation With Auxiliary Denoising Learning

3D point cloud instance segmentation, as a task in comprehending 3D scenes, involves predicting both 3D masks and class labels for individual object instances within a given point cloud. The development of an efficient transformer-based model for this task requires addressing the following key issues: refining instance masks and positions, initializing instance queries, and incorporating auxiliary task learning. To overcome the limitations of existing models, our study proposes a novel transformer-based model, T3DIS. This model refines both the mask and position, along with the query content of each instance during instance query decoding, thereby enhancing the quality of the final instance features. To expedite the instance decoding process, the model initializes the initial instance queries using a finite set of representative points selected from the point cloud. Furthermore, our approach incorporates auxiliary denoising task learning to facilitate rapid training of the transformer decoder. Through experiments conducted on the ScanNet-V2 benchmark dataset, we demonstrated the superiority of the proposed model. The evaluation involves comparing different methods of instance query initialization, position refinement, and auxiliary query denoising. © ICROS 2023.