3D Scene Segmentation using Segment Anything Model on Images Sequence with Random Prompts

3D scene segmentation suffer from low quality of reconstruction model and difficulty in semantic feature learning. Image guided methods are promising on account of the generalization ability of Segment Anything Model especially for zero-shot data. But the labels consolidation for multi-view images is still a challenge problem. It is proposed an iterative method to segment each single object with a random prompt with SAM. The method can get a good trade-off between performance and efficiency.