European Conference on Computer Vision ECCV 2026

PointLAMLocal Attentive Mamba for
Efficient Point-based 3D Object Detection

Xuanming Shang Weijia Zhang Chao Ma

Shanghai Jiao Tong UniversityMoE Key Lab of Artificial Intelligence · Institute of AI

Comparison of set-based, voxel-based, and PointLAM point-based processing, with a performance and speed plot.
Point-based geometry. Efficient detection. PointLAM preserves fine-grained point representations while offering a competitive accuracy–efficiency trade-off.

The problem & the idea

Abstract

3D object detection from LiDAR point clouds faces a core dilemma: voxel-based methods are efficient but lose geometric fidelity through quantization, while point-based methods preserve raw structure but suffer from slow downsampling and costly neighborhood queries. PointLAM addresses these bottlenecks with two complementary designs. The Laplacian Point Sampler (LPS) uses an implicit discrete Laplacian high-pass filter and Doubly Sorted Sampling to retain structure-aware foreground points efficiently. The Local Hadamard Aggregator (LHA) decouples spatial indexing from point feature representation through transient grids and replaces expensive continuous interactions with lightweight Hadamard gating. Combined with Bi-Directional Mamba layers, these components form Local Attentive Mamba blocks for efficient local and global modeling.

Inside PointLAM

A local–global view of point clouds

PointLAM architecture showing LPS, LHA, BDM-X and BDM-Y layers, BEV backbone, and detection head.
PointLAM first performs structure-aware point sampling, then stacks Local Attentive Mamba blocks that pair topology-aware local gating with axis-wise Bi-Directional Mamba layers.
01

Laplacian Point Sampler

LPS highlights high-frequency geometric structures such as boundaries and corners, then keeps salient points through fast global and regional sorting instead of iterative FPS.

02

Local Hadamard Aggregator

LHA uses a transient grid only as a deterministic spatial router. Point features remain continuous, while local topology modulates channels through Hadamard gating.

03

Local Attentive Mamba

LAM blocks anchor local geometry before serialization and use Bi-Directional Mamba layers to model long-range dependencies with linear sequence complexity.

Reported in the paper

Accuracy meets efficiency

73.0 nuScenes test NDS

68.8 mAP · single model

74.4 Waymo test L2 mAPH

79.8 L1 mAPH · single frame

90.7G FLOPs

8.6M parameters · 93.1 ms on NVIDIA A800

On nuScenes validation, PointLAM reaches 72.2 NDS and 67.8 mAP. On Waymo validation, it reaches 79.7 L1 and 73.6 L2 3D mAPH. Compared with LION, PointLAM reports higher nuScenes validation NDS with substantially fewer FLOPs and lower latency.

Seeing the details

Detection in sparse scenes

Qualitative detection examples in sparse LiDAR point clouds with ground truth and predicted boxes.
Qualitative examples show accurate localization in sparse and cluttered scenes, supporting the point-level geometric fidelity preserved by PointLAM.

What the model sees

Structure-aware features

Points retained by the sampler and processed by LHA concentrate high responses around object structures. The visualization highlights how PointLAM keeps informative foreground geometry instead of relying on coarse voxel aggregation.

Visualization of sampled points colored by feature magnitude.

Reference

Cite PointLAM

@inproceedings{shang2026pointlam,
  title={PointLAM: Local Attentive Mamba for Efficient Point-based 3D Object Detection},
  author={Shang, Xuanming and Zhang, Weijia and Ma, Chao},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}