LaSSM: Efficient Semantic-Spatial Query Decoding via Local Aggregation and State Space Models for 3D Instance Segmentation

This repository contains the weights for LaSSM, a framework for efficient 3D scene instance segmentation from point clouds.

• Paper: LaSSM: Efficient Semantic-Spatial Query Decoding via Local Aggregation and State Space Models for 3D Instance Segmentation
• Code: GitHub - RayYoh/LaSSM

Description

LaSSM (Local Aggregation and State Space Models) prioritizes simplicity and efficiency in 3D instance segmentation. It introduces a hierarchical semantic-spatial query initializer and a coordinate-guided state space model (SSM) decoder. This design restricts the model to focus on geometrically coherent regions and uses a spatial dual-path SSM block to capture underlying dependencies within the query set, reducing redundant computation.

LaSSM ranks first on the ScanNet++ V2 leaderboard, outperforming previous state-of-the-art methods with significantly fewer FLOPs.

Trained Results

Model	Benchmark	Num GPUs	mAP	AP50	AP25	Config	Tensorboard	Exp Record	Model
LaSSM	ScanNet++ V2 Val	4	29.1	43.5	51.6	Link	Link	Link	Link
LaSSM	ScanNet Val	4	58.4	78.1	86.1	Link	-	Link	Link
LaSSM	ScanNet200 Val	4	29.3	39.2	44.5	Link	-	Link	Link

Citation

@article{yao2025lassm,
  title={LaSSM: Efficient Semantic-Spatial Query Decoding via Local Aggregation and State Space Models for 3D Instance Segmentation},
  author={Yao, Lei and Wang, Yi and Yawen, Cui and Liu, Moyun and Chau, Lap-Pui},
  journal={arXiv preprint arXiv:2602.11007},
  year={2026}
}