RayMamba: Ray-Aligned Serialization for Long-Range 3D Object Detection

Cheng Lu; Mingqian Ji; Shanshan Zhang; Zhihao Li; Jian Yang

cs.CVApr 3, 2026

RayMamba: Ray-Aligned Serialization for Long-Range 3D Object Detection

Authors: Cheng Lu, Mingqian Ji, Shanshan Zhang, Zhihao Li, Jian Yang

Organizations: School of Computer Science and Technology, Nanjing University of Science and Technology, Nanjing 210094, China.

Abstract

Long-range 3D object detection remains challenging because LiDAR observations become highly sparse and fragmented in the far field, making reliable context modeling difficult for existing detectors. To address this issue, recent state space model (SSM)-based methods have improved long-range modeling efficiency. However, their effectiveness is still limited by generic serialization strategies that fail to preserve meaningful contextual neighborhoods in sparse scenes. To address this issue, we propose RayMamba, a geometry-aware plug-and-play enhancement for voxel-based 3D detectors. RayMamba organizes sparse voxels into sector-wise ordered sequences through a ray-aligned serialization strategy, which preserves directional continuity and occlusion-related context for subsequent Mamba-based modeling. It is compatible with both LiDAR-only and multimodal detectors, while introducing only modest overhead. Extensive experiments on nuScenes and Argoverse 2 demonstrate consistent improvements across strong baselines. In particular, RayMamba achieves up to 2.49 mAP and 1.59 NDS gain in the challenging 40--50 m range on nuScenes, and further improves VoxelNeXt on Argoverse 2 from 30.3 to 31.2 mAP.

Figures & tables

Fig. 1: Due to occlusion and distance-induced sparsity in LiDAR, distant objects are often represented by only a few returns.

Fig. 2: Comparison of 1D sequence context in long-range sparse scenes. For a given far-field reference voxel (red star), we highlight its context window of $K=360$ adjacent voxels in the serialized sequence. Our ray-aligned ordering (blue) preserves directionally coherent physical structures, whereas the Hilbert ordering (purple) activates spatially scattered, unrelated regions.

Fig. 3: Overview of RayMamba. Top: RayMamba blocks are inserted into a sparse 3D convolutional backbone. Bottom: Structure of a RayMamba block. RayMamba consists of two components: Ray-Aligned Serialization , which converts sparse voxel features into sector-wise ordered sequences using an offline-generated dense sector template, and SectorMamba3D , which performs sector-wise sequence modeling before the enhanced features are restored to sparse 3D space through Sequence-to-Spatial and sparse deconvolution.

Fig. 4: Ray-aligned serialization strategy. (a) Azimuth sector partitioning: The BEV space is divided into independent angular sectors to separate directionally distinct regions. (b) Sector-wise ordering: Voxels in each sector are serialized by first traversing height layers from top to bottom, introducing a vertical layering prior, and then applying angular ordering within each layer to preserve directional continuity. For a fixed voxel grid, the resulting sector assignments and ordering scores are precomputed as a dense sector template, which is queried at runtime to convert active sparse voxels into sector-wise ordered sequences.

Method	Data	mAP	NDS	Car	Truck	C.V	Bus	Trailer	Barrier	Motor.	Bike	Ped.	T.C.
PointPillars [ 33 ]	L	34.3	49.1	79.0	38.7	5.9	52.1	25.0	42.4	21.3	0.8	60.0	18.0
SSN [ 34 ]	L	46.7	58.2	83.0	52.1	16.2	61.7	32.1	54.1	47.9	22.3	70.1	27.2
CenterPoint [ 14 ]	L	56.4	64.7	84.6	53.9	16.5	67.0	37.5	65.0	55.3	36.7	83.3	64.3
+ RayMamba	L	57.4	65.6	85.5	54.4	16.7	68.2	38.2	63.5	58.7	38.5	84.1	66.6
TransFusion [ 35 ]	LC	67.5	71.3	-	-	-	-	-	-	-	-	-	-
BEVFusion [ 36 ]	LC	69.6	72.1	89.1	66.7	30.9	77.7	42.6	73.5	79.0	67.5	89.4	79.3

TABLE I: Performance comparison on the nuScenes validation set. ’L’ denotes LiDAR-only methods, and ’LC’ denotes LiDAR-Camera multi-modal methods. The best results within each modality group are highlighted in bold .

Method	Overall		0–20m		20–40m		40–50m
Method	mAP	NDS	mAP	NDS	mAP	NDS	mAP	NDS
CenterPoint	56.4	64.7	72.3	74.0	46.0	58.6	26.7	53.4
+RayMamba ( $11\times 256^{2}$ )	57.4	65.6	73.2	75.1	47.5	59.6	29.2	55.0
vs. baseline	+1.0	+0.9	+0.9	+1.1	+1.5	+1.0	+2.5	+1.6
+RayMamba ( $21\times 512^{2}$ )	57.7	65.9	73.2	75.1	48.1	60.3	27.2	54.0
vs. baseline	+1.3	+1.2	+0.9	+1.1	+2.1	+1.7	+0.5	+0.6

TABLE II: Performance comparison of RayMamba across different scales relative to CenterPoint baseline on nuScenes.

Method	Overall		0–20m		20–40m		40–50m
Method	mAP	NDS	mAP	NDS	mAP	NDS	mAP	NDS
MV2DFusion	73.0	74.8	82.6	80.7	66.0	70.4	43.0	63.0
+RayMamba ( $30^{\circ}$ )	73.2	74.8	82.6	80.5	66.3	70.7	43.3	63.1
vs. baseline	+0.1	+0.1	+0.0	-0.2	+0.3	+0.3	+0.3	+0.1
+RayMamba ( $60^{\circ}$ )	73.4	75.0	83.0	80.9	66.3	70.9	44.6	63.7
vs. baseline	+0.3	+0.3	+0.4	+0.2	+0.3	+0.5	+1.6	+0.7

TABLE III: Ablation study of the sector angular step

\Delta\theta

in RayMamba on the MV2DFusion baseline on the nuScenes validation set. Best results are highlighted in bold.

Fig. 5: Qualitative comparison on challenging long-range occluded targets. Green boxes denote ground truth, red dashed boxes denote baseline predictions, and blue dashed boxes denote RayMamba predictions.

Method	Overall		0–20m		20–40m		40–50m
Method	mAP	NDS	mAP	NDS	mAP	NDS	mAP	NDS
MV2DFusion + RayMamba	73.4	75.0	83.0	80.9	66.3	70.9	44.6	63.7
+ distance sort	72.9	74.7	82.7	80.9	65.3	70.0	44.7	63.8

TABLE IV: Ablation study of explicit distance sorting in RayMamba on the MV2DFusion baseline on the nuScenes validation set. Best results are highlighted in bold.

Model	mAP 40-50m $\uparrow$	NDS 40-50m $\uparrow$	Params (M) $\downarrow$	FPS $\uparrow$
Base Architecture: MV2DFusion (Multi-modal)
Baseline	42.99	63.02	101.27	1.05
+ RayMamba	44.59	63.70	103.74	1.01
Improvement	+1.60	+0.68	+2.47	-0.04
Base Architecture: CenterPoint (LiDAR-only)
Baseline	26.74	53.44	8.94	11.20

TABLE V: Efficiency and performance comparison on the nuScenes validation set. All metrics are evaluated in the 40–50 m range.

Methods	Average	Vehicle	Bus	Pedestrian	Stop Sign	Box Truck	Bollard	C-Barrel	Motorcyclist	MPC-Sign	Motorcycle	Bicycle	A-Bus	School Bus	Truck Cab	C-Cone	V-Trailer	Sign	Large Vehicle	Stroller	Bicyclist
CenterPoint [ 14 ]	22.0	67.6	38.9	46.5	16.9	37.4	40.1	32.2	28.6	27.4	33.4	24.5	8.7	25.8	22.6	29.5	22.4	6.3	3.9	0.5	20.1
FSD [ 27 ]	28.2	68.1	40.9	59.0	29.0	38.5	41.8	42.6	39.7	26.2	49.0	38.6	20.4	30.5	14.8	41.2	26.9	11.9	5.9	13.8	33.4
VoxelNeXt [ 4 ]	30.3	72.1	38.5	63.2	37.5	40.5	52.8	64.9	44.4	35.3	44.8	39.3	21.5	27.5	17.4	45.4	21.2	16.4	7.1	12.3	33.3
+ RayMamba	31.2	73.7	39.8	64.6	40.1	40.5	56.5	69.3	46.0	34.5	47.3	40.9	20.6	26.4	19.6	45.5	21.5	16.6	6.1	15.7	35.0

TABLE VI: Further validation on the AV2validation set. RayMamba is applied to the VoxelNeXt baseline to examine its generalization in more extreme long-range sparse scenes.

Method	Overall		0–30m		30–50m		50–150m
Method	mAP	CDS	mAP	CDS	mAP	CDS	mAP	CDS
VoxelNext	30.3	22.9	48.7	38.5	34.2	26.3	16.5	11.7
+RayMamba ( $13\times 250^{2}$ )	30.2	22.8	48.0	38.1	33.3	25.6	17.3	12.3
vs. baseline	-0.1	-0.1	-0.7	-0.4	-0.9	-0.7	+0.8	+0.6
+RayMamba ( $26\times 500^{2}$ )	30.6	23.1	48.2	38.2	34.4	26.4	17.4	12.4
vs. baseline	+0.3	+0.2	-0.5	-0.3	+0.2	+0.1	+0.9	+0.7

TABLE VII: Ablation study of RayMamba template sizes on the VoxelNeXt baseline on the AV2validation set. Best results are highlighted in bold.

Method	Overall		0–30m		30–50m		50–150m
Method	mAP	CDS	mAP	CDS	mAP	CDS	mAP	CDS
VoxelNeXt + RayMamba	31.2	23.7	49.5	39.3	33.7	26.0	18.0	12.8
VoxelNeXt + HilbertMamba	30.6	23.1	48.3	38.2	33.7	25.9	17.5	12.5

TABLE VIII: Comparison of sequence serialization strategies on the AV2 validation set based on the VoxelNeXt baseline. RayMamba and HilbertMamba use the same template-size configuration for a fair comparison. Best results are highlighted in bold.

3D Object Detection

3D object detection identifies objects and estimates their locations, sizes, and orientations from images, point clouds, radar, or fused sensors.

See all 180 papers

PointLAM: Local Attentive Mamba for Efficient Point-based 3D Object Detection
Sep 18, 2026Xuanming Shang, Weijia Zhang, Chao Ma3D Object Detection Efficient Object Detection
MoE Key Lab of Artificial Intelligence, Institute of AI, Shanghai Jiao Tong University, Shanghai, China
Fore-Mamba3D: Mamba-based Foreground-Enhanced Encoding for 3D Object Detection
Feb 23, 2026Zhiwei Ning, Xuanang Gao, Jiaxi Cao +53D Object Detection Mamba
School of Automation and Intelligent Sensing, Shanghai Jiao Tong University · Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University · School of Computer Science and Technology, Zhejiang Normal University +2
ATN3D: Density-Aware LiDAR-Radar Early 3D Object Detection Under Extreme Sparsity
Jun 8, 2026Debojyoti Biswas, Xianbiao HuLidar-Based 3D Object Detection Autonomous Driving Perception

Sep 18, 2026J/K move · Enter open · S save

PointLAM: Local Attentive Mamba for Efficient Point-based 3D Object Detection

Xuanming Shang, Weijia Zhang, Chao Ma

MoE Key Lab of Artificial Intelligence, Institute of AI, Shanghai Jiao Tong University, Shanghai, China

3D Object Detection Efficient Object Detection

Open paper AskPDF arXiv

3D object detection from LiDAR point clouds faces a fundamental dilemma: voxel-based methods achieve efficiency at the cost of geometric quantization, while point-based methods preserve fidelity but suffer from prohibitive computational bottlenecks. Specifically, point-based architectures are crippled by slow downsampling strategies (e.g., FPS) and expensive dynamic neighbor queries (e.g., k-NN) coupled with costly continuous interactions. To tackle these systemic inefficiencies, we propose PointLAM, a highly efficient and powerful point-based architecture driven by two synergistic innovations. First, to resolve the downsampling bottleneck, we develop the Laplacian Point Sampler (LPS). LPS employs an implicit discrete Laplacian high-pass filter and Doubly Sorted Sampling to achieve fast, structure-aware foreground preservation. Second, to overcome local modeling latency, we design the Local Hadamard Aggregator (LHA). LHA decouples spatial indexing from feature representation using transient grids, and replaces complex continuous interactions with a Hadamard Gating mechanism for topology-aware, attentive modulation. By coupling this local gating with Bi-Directional Mamba (BDM) layers for global sequence modeling, we formulate the Local Attentive Mamba (LAM) block. Powered by this architecture, PointLAM achieves competitive performance on nuScenes and Waymo for point-based detectors. It rivals highly optimized voxel competitors while requiring a fraction of the computational footprint, demonstrating marked superiority in detecting small instances and handling extreme sparsity. Project page: https://pointlam.github.io/.

Abstract

3D Object Detection

Explore similar work

PointLAM: Local Attentive Mamba for Efficient Point-based 3D Object Detection

Fore-Mamba3D: Mamba-based Foreground-Enhanced Encoding for 3D Object Detection

ATN3D: Density-Aware LiDAR-Radar Early 3D Object Detection Under Extreme Sparsity