3D-SDIS：Enhanced 3D Instance Segmentation through Spectral Fusion and Dual-Sphere Sampling

3D instance segmentation is essential in applications such as autonomous driving, augmented reality, and robotics, where accurate identification of individual objects in complex point cloud data is required. Existing methods typically rely on feature learning in a single spatial domain and often fail in cases involving overlapping objects and sparse point distributions. To solve these problems, we propose 3D-SDIS, a multi-domain 3D instance segmentation network. It includes an FFT-Spatial Fusion Encoder (FSF Encoder) that uses the Fourier transform to convert spatial features into the frequency domain. This process reduces interference from redundant points and improves boundary localization. We also introduce an Offset Dual-Sphere Sampling Module (ODSS), which performs multi-view feature sampling based on both the original and offset sphere centers. It increases the receptive field and captures more geometric information. Experimental results on the ScanNetV2 (62.9) and S3DIS (61.0) datasets demonstrate the superiority of 3D-SDIS over state-of-the-art methods, particularly in handling overlapping instances and large planar structures. The source code and trained models are available at http://github.com/cbbbbg/3D-SDIS.