SRMambaV2: LiDAR Sparse Upsampling Method

• SRMambaV2 is a sparse point cloud upsampling method that reconstructs dense 3D data from sparse LiDAR scans by reformulating the task as 2D range-image super-resolution.
• It employs a biomimetic 2D selective scanning self-attention mechanism and a dual-branch U-shaped network to enhance sparse feature recovery and geometric fidelity.
• A progressive adaptive loss enforces bird’s-eye-view consistency, leading to improved metrics on KITTI-360 and nuScenes compared to previous methods.

Searching arXiv for SRMambaV2 and closely related Mamba-based super-resolution papers to ground the article with current citations. SRMambaV2 is a sparse point cloud upsampling method for autonomous driving that reconstructs denser 3D point clouds from sparse LiDAR scans by reformulating the task as a 2D range-image super-resolution problem. It is presented as a hybrid scan-plus-attention framework centered on a biomimetic 2D selective scanning self-attention mechanism, a dual-branch U-shaped architecture, and a progressive adaptive loss designed to emphasize sparse distant regions while enforcing bird’s-eye-view geometric consistency (Chen et al., 23 Jul 2025). Within the broader Mamba-based super-resolution literature, it occupies a distinct position: unlike remote-sensing image SR systems such as FMSR or lightweight image SR systems such as DVMSR, SRMambaV2 addresses LiDAR-specific sparsity, projection geometry, and long-range degradation in autonomous driving (Xiao et al., 2024, Lei et al., 2024).

1. Problem setting and motivation

SRMambaV2 addresses the problem of recovering dense, geometrically accurate 3D point clouds from sparse LiDAR scans, with particular emphasis on far-range regions where the data is extremely sparse and noisy (Chen et al., 23 Jul 2025). The method converts 3D point cloud upsampling into a 2D range-image super-resolution task, where each pixel stores depth-related information from the LiDAR projection. This reformulation makes the problem computationally manageable, but it also creates a representation in which information density is highly non-uniform: near objects occupy many pixels, whereas far objects appear as weak, blurry, sparse traces (Chen et al., 23 Jul 2025).

The paper attributes the difficulty of prior range-image methods to several factors. It identifies depth-dependent sparsity, blurry sparse features, and the tendency of bottom-up feature hierarchies to discard the very details needed for point recovery. It also argues that local-window attention is insufficient for modeling global structure across sparse regions, and that standard image SR losses are not geometry-aware, so directly applying image super-resolution methods to range images can produce significant noise and poor 3D geometric fidelity (Chen et al., 23 Jul 2025).

The formal objective is defined from a sparse point cloud

$$P_{LR} = \{p_1, p_2, \ldots, p_n\}, \quad p_i = \{x_i, y_i, z_i\},$$

to a denser point cloud

$$P_{HR} = \{p_1, p_2, \ldots, p_n, \ldots, p_{4n}\}.$$

In the range-image formulation, the low-resolution input has size $H \times W$ and the target high-resolution output has size $4H \times W$, corresponding to 4× upsampling along the vertical beam dimension (Chen et al., 23 Jul 2025).

A plausible implication is that SRMambaV2 should not be understood as a generic Mamba SR model transplanted to 3D data. Its design is tied to the statistical and geometric properties of projected LiDAR range images, especially anisotropic continuity and sparse long-range structure.

2. Scan-to-focus formulation and overall architecture

The method is explicitly inspired by human driver visual perception. The paper describes a Scan-to-Focus strategy with three stages: scanning, modulation, and focusing (Chen et al., 23 Jul 2025). Scanning collects global structural cues from all directions, modulation recalibrates features and adjusts receptive field size, and focusing uses transformer attention to refine salient structures. This strategy is implemented in a U-Net-like encoder-decoder with skip connections (Chen et al., 23 Jul 2025).

Before the main network, the method applies Hough voting and depth inpainting, following the earlier SRMamba pipeline. It also uses horizontal 1D convolution to encode lateral information, producing a latent feature

$I_{\text{latent} \in \mathbb{R}^{H \times \frac{W}{4} \times C_1}.$

The paper motivates this by noting that horizontal continuity in range images is stronger than vertical continuity (Chen et al., 23 Jul 2025).

The network combines three principal components:

Component	Role	Stated function
2DSSA	Encoder and decoder scanning	Models feature distribution in distant sparse areas
Dual-branch modulation	Intermediate recalibration	Enhances sparse feature representation
PAL	Training objective	Refines fine-grained details during upsampling

The architecture is therefore neither a pure state-space model nor a pure transformer. The paper characterizes it as a hybrid design in which SS2D provides a coarse global prior and Swin Transformer blocks provide refinement (Chen et al., 23 Jul 2025). This distinguishes SRMambaV2 from image SR models such as DVMSR, which use stacked Residual State Space Blocks with unidirectional Vision Mamba Modules and a distillation strategy (Lei et al., 2024), and from remote-sensing SR models such as FMSR, which combine VSSM with frequency selection and local gating (Xiao et al., 2024).

3. Biomimetic 2D selective scanning self-attention

The central novelty of SRMambaV2 is the biomimetic 2D selective scanning self-attention mechanism, abbreviated 2DSSA (Chen et al., 23 Jul 2025). The mechanism is intended to imitate a human driver’s perception by first performing coarse scanning of the scene, then extracting a global structural prior, and finally focusing on important sparse regions.

For the $i$-th block, the input feature is

$F_{in} \in \mathbb{R}^{c \times h \times w}.$

It is normalized, processed through SS2D, and combined through a residual path: $F_{scan} = Residual(SS2D(Norm(F_{in})), F_{in}).$ The SS2D module scans in four directions: left-to-right, right-to-left, top-to-bottom, and bottom-to-top (Chen et al., 23 Jul 2025). The stated purpose is to extract global contextual information while retaining 2D spatial structure.

This directional scanning is a key design choice. Instead of flattening the range image into a single sequence, the method preserves directional structure and builds what the paper describes as a coarse global overview, particularly for sparse long-range regions where local features are weak (Chen et al., 23 Jul 2025). The paper further argues that standard attention may over-focus on strong dense signals and miss weak but important sparse structures; the scanning stage provides a global feedback signal intended to counteract that bias (Chen et al., 23 Jul 2025).

This design can be situated relative to other Mamba-based SR systems. FMSR uses a Vision State Space Module to capture long-range dependencies with linear complexity in remote sensing image SR, but supplements it with frequency selection and local gating because pure spatial long-range modeling is insufficient for recovering fine details (Xiao et al., 2024). DVMSR uses unidirectional SSM inside a Vision Mamba Module for efficient image SR, emphasizing low parameter count and distillation rather than sparse-region modeling (Lei et al., 2024). SRMambaV2 adapts the state-space idea to directional 2D scanning over range images and explicitly ties it to sparse LiDAR geometry (Chen et al., 23 Jul 2025).

4. Modulation and focus stages

After scanning, SRMambaV2 applies a modulation stage designed to refine channel-wise semantics and control receptive field size (Chen et al., 23 Jul 2025). The modulation block has a dual-branch structure, with a downsampling branch in the encoder and an upsampling branch in the decoder. A modified SE layer is used in which global average pooling is replaced by global max pooling to preserve salient sparse activations.

The paper gives the formulation as

$$\left\{ \begin{matrix} F_{mid} = Residual(SE(Norm(F_{scan})), F_{scan}) \ F_{down} = Norm(Conv2d(F_{mid})) \ F_{up} = PixelShuffle(Conv2d(F_{mid})) \end{matrix} \right.$$

where $F_{mid}$ is the recalibrated feature,

$$P_{HR} = \{p_1, p_2, \ldots, p_n, \ldots, p_{4n}\}.$$0

and

$$P_{HR} = \{p_1, p_2, \ldots, p_n, \ldots, p_{4n}\}.$$1

According to the paper, the downsampling branch increases receptive field in the encoder, while the upsampling branch restores spatial detail in the decoder (Chen et al., 23 Jul 2025).

The focus stage uses Swin Transformer blocks. The paper specifies W-MSA and SW-MSA, with patch size set to $$P_{HR} = \{p_1, p_2, \ldots, p_n, \ldots, p_{4n}\}.$$2 to suit range-image geometry (Chen et al., 23 Jul 2025). The output of the modulation stage is injected as a context prior, guiding attention toward salient regions identified during scanning. The stated roles of this stage are to refine boundaries, sharpen geometric details, capture cross-window dependencies, and integrate globally informed coarse scanning with local fine attention (Chen et al., 23 Jul 2025).

This hybrid decomposition into scan, modulation, and focus is central to the method’s identity. It suggests that SRMambaV2 treats sparse-region reconstruction as a problem of staged feature control rather than relying on a single homogeneous backbone. A plausible implication is that the method inherits some of the efficiency and long-range modeling motivation associated with state-space scanning, while retaining transformer-based refinement where localized detail integration is required.

5. Progressive adaptive loss and geometric supervision

The training objective is a progressive adaptive loss, abbreviated PAL. The paper argues that standard $$P_{HR} = \{p_1, p_2, \ldots, p_n, \ldots, p_{4n}\}.$$3 loss is insufficient because it treats all pixels equally, whereas the importance of range-image pixels is highly non-uniform (Chen et al., 23 Jul 2025). The total loss is defined as

$$P_{HR} = \{p_1, p_2, \ldots, p_n, \ldots, p_{4n}\}.$$4

The base reconstruction term is

$$P_{HR} = \{p_1, p_2, \ldots, p_n, \ldots, p_{4n}\}.$$5

The adaptive weighted term is

$$P_{HR} = \{p_1, p_2, \ldots, p_n, \ldots, p_{4n}\}.$$6

where $$P_{HR} = \{p_1, p_2, \ldots, p_n, \ldots, p_{4n}\}.$$7 is a learnable weighting mask emphasizing sparse but structurally critical areas, especially distant regions (Chen et al., 23 Jul 2025).

The geometric term is the BEV consistency loss: $$P_{HR} = \{p_1, p_2, \ldots, p_n, \ldots, p_{4n}\}.$$8 The paper states that this term supervises the reconstructed point cloud in bird’s-eye view and helps reduce projection-induced 3D misalignment and ghost points (Chen et al., 23 Jul 2025).

This loss construction is one of the most important distinctions between SRMambaV2 and image-domain Mamba SR methods. DVMSR uses $$P_{HR} = \{p_1, p_2, \ldots, p_n, \ldots, p_{4n}\}.$$9 and an output-feature distillation term for efficient image SR (Lei et al., 2024), while FMSR emphasizes architecture-level spatial-frequency fusion for remote sensing image SR (Xiao et al., 2024). SRMambaV2, by contrast, integrates region-aware weighting and BEV consistency because range-image fidelity alone does not guarantee accurate 3D structure (Chen et al., 23 Jul 2025).

The paper also notes that introducing PAL later in training further improves performance, which it interprets as evidence that PAL functions as a refinement-oriented objective (Chen et al., 23 Jul 2025).

6. Projection model, datasets, metrics, and quantitative results

The appendix provides the spherical projection from a 3D point $H \times W$0 to range-image coordinates: $H \times W$1 where $H \times W$2 and $H \times W$3 are beam-related arrays. Hough voting is used to reduce truncation errors during projection (Chen et al., 23 Jul 2025).

The method is evaluated on KITTI-360 and nuScenes. KITTI-360 uses 20,000 training scans and 2,500 validation scans; nuScenes uses 28,130 training scans and 6,008 validation scans. Both are 360° outdoor driving datasets with Velodyne LiDAR (Chen et al., 23 Jul 2025). The experiments simulate sparse inputs by downsampling point clouds by 4×. Range-image sizes are $H \times W$4 and $H \times W$5 for input, and $H \times W$6 and $H \times W$7 for output (Chen et al., 23 Jul 2025).

Training uses 1000 epochs, checkpointing every 50 epochs, best-model selection by validation IoU, the AdamW optimizer, and an initial learning rate of 0.005. Reported hardware is 4× NVIDIA Tesla V100-PCIE-16GB, Intel Xeon Silver 4210, and 16 GB RAM (Chen et al., 23 Jul 2025). Evaluation metrics are IoU, Chamfer Distance, Jensen-Shannon Divergence, and MAE (Chen et al., 23 Jul 2025).

The reported quantitative results are as follows:

Dataset	Method	IoU	CD	MAE	JSD
KITTI-360	SRMamba	0.4389	0.1031	0.0044	0.0052
KITTI-360	SRMambaV2	0.4516	0.0826	0.0041	0.0045
nuScenes	SRMamba	0.3170	1.0196	0.0287	0.0293
nuScenes	SRMambaV2	0.3299	0.9485	0.0284	0.0267

Compared with TULIP, the paper reports on KITTI-360: +8.7% IoU, 33.4% lower CD, 19.6% lower MAE, and 35.7% lower JSD. Compared with SRMamba, it reports +2.9% IoU and 19.9% CD reduction. On nuScenes, compared with TULIP it reports +8.2% IoU, 9.7% lower CD, 3.1% lower MAE, and 12.2% lower JSD; compared with SRMamba, it reports +4.1% IoU and 7.0% CD reduction (Chen et al., 23 Jul 2025).

The paper emphasizes that SRMambaV2 achieves particularly strong gains in Chamfer Distance, which it interprets as improved geometric fidelity (Chen et al., 23 Jul 2025). Qualitatively, it reports cleaner object contours, fewer spurious points, better recovery of sparse vehicle shapes, and stronger robustness in long-range and cluttered scenes (Chen et al., 23 Jul 2025).

7. Ablation findings, limitations, and relationship to related Mamba SR methods

The ablation study presents cumulative improvements from each component. On KITTI-360, baseline SRMamba reports IoU 0.4389, CD 0.1031, and MAE 0.0044. Adding 2DSSA yields IoU 0.4416, CD 0.0869, and MAE 0.0043. Adding 2DSSA plus Modulation gives IoU 0.4468, CD 0.0864, and MAE 0.0044. Adding all components, including PAL, gives IoU 0.4516, CD 0.0826, and MAE 0.0041 (Chen et al., 23 Jul 2025).

On nuScenes, baseline SRMamba reports IoU 0.3170, CD 1.0196, and MAE 0.0287. Adding 2DSSA yields IoU 0.3199, CD 0.9616, and MAE 0.0294. Adding 2DSSA plus Modulation gives IoU 0.3210, CD 0.9718, and MAE 0.0293. The full model gives IoU 0.3299, CD 0.9485, and MAE 0.0284 (Chen et al., 23 Jul 2025). The paper interprets these results by stating that 2DSSA contributes the biggest gain in global structure and CD, Modulation helps feature recalibration and detail recovery, and PAL provides the final boost, especially in sparse and geometrically difficult regions (Chen et al., 23 Jul 2025).

The main reported limitation is a failure case in noisy scenes with irregular vegetation, where the method may still misinterpret scattered structures as continuous contours (Chen et al., 23 Jul 2025). The paper attributes this to the fact that range images still lack explicit 3D structural context. It also implies other practical constraints: dependence on the range-image projection pipeline, reliance on preprocessing such as Hough voting and depth inpainting, and greater architectural complexity than a plain CNN baseline (Chen et al., 23 Jul 2025).

Relative to other Mamba-based super-resolution methods, SRMambaV2 belongs to a shared trend of replacing or augmenting CNN/Transformer backbones with state-space scanning for long-range dependency modeling, but its problem domain and technical choices are substantially different. FMSR is a remote sensing image super-resolution framework that uses Vision State Space Modules together with Frequency Selection Module, Hybrid Gate Module, and learnable scaling adaptors to fuse global spatial dependency, frequency-domain cues, and local spatial detail (Xiao et al., 2024). DVMSR is a lightweight image super-resolution network built from Residual State Space Blocks with unidirectional Vision Mamba Modules and trained with teacher-student distillation (Lei et al., 2024). SRMambaV2 instead targets automotive sparse point cloud upsampling through range-image processing, biomimetic directional scanning, dual-branch modulation, and geometry-aware supervision (Chen et al., 23 Jul 2025).

This suggests that “SRMambaV2” should not be conflated with Mamba-based image SR in the conventional sense. Its central contribution is not merely the use of state-space modeling, but the integration of selective 2D scanning, sparse-region-aware modulation, and BEV-consistent loss within a LiDAR-specific upsampling pipeline (Chen et al., 23 Jul 2025).