Multimodal Plane Instance Segmentation with the Segment Anything Model

Plane instance segmentation from RGB-D data is critical for BIM-related tasks. However, existing deep-learning methods rely on only RGB bands, overlooking depth information. To address this, PlaneSAM, a Segment-Anything-Model-based network, is proposed. It fully integrates RGB-D bands using a dual-complexity backbone: a simple branch primarily for the D band and a high-capacity branch mainly for RGB bands. This structure facilitates effective D-band learning with limited data, preserves EfficientSAM's RGB feature representations, and enables task-specific fine-tuning. To improve adaptability to RGB-D domains, a self-supervised pretraining strategy is introduced. EfficientSAM's loss is also optimized for large-plane segmentation. Additionally, plane detection is performed using Faster R-CNN, enabling fully automatic segmentation. State-of-the-art performance is achieved on multiple datasets, with <10% additional overhead compared to EfficientSAM. The proposed dual-complexity backbone shows strong potential for transferring RGB-based foundation models to RGB+X domains in other scenarios, while the pretraining strategy is promising for other data-scarce tasks.