Instance Position Encoding and Dual Denoising Tasks for Efficient 3D Instance Segmentation

As a 3D scene understanding task, 3D point cloud instance segmentation is to predict simultaneously both their 3D mask and class label of individual object instances included in a given point cloud. There are some important issues to design an efficient Transformer-based 3D point cloud instance segmentation model: instance query position encoding, instance mask and query refinement, auxiliary task learning. In order to overcome the limitations of existing models, this paper proposes a novel Transformer-based model, DP3DIS, which contains some distinct solutions such as position encoding based on 3D bounding box of an instance, a new masked modulate cross attention with consideration of instance size or volume, denoising auxiliary tasks applied for both instance query and instance attention mask. In this paper, we demonstrate the superiority of the proposed model Through quantitative and qualitative experiments conducted on ScanNet v2, we showed the superiority of the proposed DP3DIS, model.