Biomimetic 2DSSA for LiDAR Upsampling

• Biomimetic 2DSSA is a hybrid mechanism that leverages multi-directional 2D scanning and focused self-attention to improve LiDAR range image upsampling.
• The method enhances distant, sparse region reconstruction by integrating a global SS2D scan with local Swin Transformer refinement.
• Its biomimetic, coarse-to-fine strategy emulates human visual attention, effectively mitigating depth-dependent spatial imbalances in autonomous driving data.

Biomimetic 2D Selective Scanning Self-Attention (2DSSA) is a hybrid scan-to-focus feature-processing mechanism introduced in SRMambaV2 for range-image-based LiDAR point cloud upsampling in autonomous driving. It is designed to improve reconstruction in distant, sparse regions by combining 2D-Selective-Scan (SS2D) for global contextual browsing with Swin Transformer windowed self-attention for focused refinement, under a biomimetic analogy to human driver vision that first performs a rough global scan and then concentrates on salient distant targets (Chen et al., 23 Jul 2025).

1. Problem setting and motivation

2DSSA is motivated by a specific failure mode in range-image super-resolution for LiDAR. Recent methods convert 3D point cloud upsampling into 2D range-image super-resolution, but “due to the sparse and blurry feature representation of range images, accurately reconstructing detailed and complex spatial topologies remains a major difficulty” (Chen et al., 23 Jul 2025). The underlying issue is not only sparsity in the point cloud itself, but also the structured nonuniformity of the projected range image.

The method is framed around what the SRMambaV2 authors describe as a depth-dependent spatial imbalance. As distance increases, LiDAR returns become sparser, object structure becomes more incomplete, contours blur, and structurally important but weak signals are easier to miss. The paper states that “range images exhibit significant variations in information density across different depth regions,” and that this “depth-dependent spatial imbalance substantially increases the difficulty for models to perform structural perception in distant areas” (Chen et al., 23 Jul 2025). It further notes that upper rows of the range image often correspond to distant regions and contain sparse but concentrated structure.

A second motivation is directed at local attention. The paper explicitly criticizes window-constrained attention on the grounds that “attention mechanisms constrained by local windows struggle to capture structures beyond individual windows, exhibiting limited global modeling capability” (Chen et al., 23 Jul 2025). Within this framing, 2DSSA is intended to provide global perception before local refinement rather than relying only on bottom-up hierarchical extraction.

The biomimetic argument is central to the module’s justification. The authors ground the design in human visual attention during driving, which they describe as following a coarse-to-fine processing strategy: the system first forms a rough overall perception and then prioritizes salient or task-critical areas, especially distant ones. In that sense, 2DSSA is not merely a sparse attention block; it is presented as a top-down mechanism for overview-first contextual guidance.

2. Architectural composition inside SRMambaV2

Within SRMambaV2, 2DSSA is the basic feature-processing unit used in both encoder and decoder stages of a U-Net-like dual-branch architecture. The architecture description states that “The method uses a U-Net architecture, employing 2DSSA as the encoder and decoder, which comprises three main steps: (a) scanning, (b) modulation, and (c) focus” (Chen et al., 23 Jul 2025). This establishes that 2DSSA is an integrated module rather than a single isolated attention layer.

The module begins from an input feature map

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

After scanning, the output remains

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

In the modulation stage, two branch-specific transforms are defined. The downsampling branch produces

$$F_{down}\in R^{2c\times \frac{h}{2} \times \frac{w}{2}},$$

while the upsampling branch produces

$F_{up}\in R^{\frac{c}{2}\times 2h \times 2w}.$

The paper also describes an earlier latent representation $I_{latent}$ of dimension

$H\times \frac{W}{4} \times C_{1},$

obtained after horizontal 1D convolution encoding of the optimized range image. The manuscript mixes $H\times W\times C$ and $C\times H\times W$ notation, so the intended tensor semantics are clearer than the notation is standardized (Chen et al., 23 Jul 2025).

The scanning stage applies SS2D after normalization. The paper states that “Selective scanning is performed in four directions—left to right, right to left, top to bottom, and bottom to top—to achieve global perception and highlight feature-rich regions” (Chen et al., 23 Jul 2025). This four-direction traversal is the most explicit algorithmic description of the 2D scanning process in the paper.

The modulation stage applies an SE-like recalibration block and then a branch-specific resize operator. A notable implementation detail is that the SELayer is modified by replacing global average pooling with global max pooling “to enhance its adaptability to salient feature regions” (Chen et al., 23 Jul 2025). In the encoder, a strided convolution downsamples and doubles channels; in the decoder, a $1\times1$ convolution plus PixelShuffle upsamples and rearranges channels.

The focus stage uses Swin Transformer refinement. The paper states that the modulated feature is used “as a context prior to guide the attention mechanism,” and that the feature map is partitioned into non-overlapping patches of size $2\times 8$, followed by W-MSA and SW-MSA (Chen et al., 23 Jul 2025). This is why the module is best understood as hybrid: SS2D performs the global overview, while Swin provides subsequent local-to-meso-scale attention refinement.

3. Mathematical formulation and explicitly stated operations

The paper gives only a small number of explicit equations for 2DSSA itself. The scanning stage is defined by

$F_{scan}\in R^{c\times h \times w}.$0

This equation specifies that the input feature is normalized, processed by SS2D, and then added back to the original feature through a residual connection (Chen et al., 23 Jul 2025).

The modulation stage is given as

$F_{scan}\in R^{c\times h \times w}.$1

where $F_{scan}\in R^{c\times h \times w}.$2 is the modulated intermediate feature, $F_{scan}\in R^{c\times h \times w}.$3 is the downsampled encoder output, and $F_{scan}\in R^{c\times h \times w}.$4 is the upsampled decoder output (Chen et al., 23 Jul 2025). The manuscript contains an apparent typographical issue in the $F_{scan}\in R^{c\times h \times w}.$5 expression, which includes an extra right parenthesis.

These equations establish the residual placement, the presence of normalization before SS2D and SE, and the branch-specific transforms. They do not formalize the internals of SS2D, the directional fusion rule across the four scans, or the internal attention equations of the Swin focus stage. The paper also does not provide explicit formulas for query, key, value projections, shifted-window masking, attention score computation, selective gate coefficients, or computational complexity for 2DSSA itself (Chen et al., 23 Jul 2025).

Accordingly, the mathematically precise core of 2DSSA, as documented in the paper, is limited to the residualized scan stage and the residualized modulation stage. A plausible implication is that the authors treat SS2D as an imported primitive from VMamba-style visual state-space modeling and use Swin as an imported primitive for focused attention, while defining the novelty primarily at the module-composition level rather than through a new closed-form attention operator.

4. Biomimetic interpretation

The biomimetic designation is tied to a functional analogy rather than to a neuroscientific formalization. The authors explicitly claim that human visual attention “typically follows a coarse-to-fine processing strategy, where the brain first performs a rough scan to quickly form an overall high-level perception, which is then integrated with sensory input” (Chen et al., 23 Jul 2025). For autonomous driving, they further argue that “the visual attention of the driver typically prioritizes key information in distant regions, which aligns with the processing strategy of ‘coarse perception followed by fine attention’” (Chen et al., 23 Jul 2025).

In architectural terms, the analogy maps directly onto the three-stage organization of 2DSSA. The SS2D scan supplies a rough global overview. The modulation stage injects feedback-like feature recalibration. The Swin focus stage then allocates attention to salient structure under the guidance of the prior context. The paper consistently contrasts this with purely bottom-up feature extraction, which it argues can lose spatial detail too early.

The paper does not use the terms fovea, foveation, or peripheral-to-central processing. Any interpretation of 2DSSA as a foveal-like mechanism is therefore an inference rather than a direct claim. What is directly supported is overview-first processing, feedback-guided focus, and preferential treatment of distant sparse regions. The biomimetic content is therefore best described as a coarse-to-fine, top-down visual strategy specialized to the geometry of range images.

This biomimetic framing also explains why the module is not reduced to local window attention alone. The overview phase is intended to identify structurally weak but globally meaningful distant signals before they are refined by attention. In the paper’s own terms, this is meant to prevent the model from overemphasizing strong signals while overlooking weak but structurally important ones.

5. Empirical behavior and reported performance

The clearest isolated evidence for 2DSSA comes from ablation results on KITTI-360 and nuScenes. On KITTI-360, the baseline reports IoU $F_{scan}\in R^{c\times h \times w}.$6, CD $F_{scan}\in R^{c\times h \times w}.$7, and MAE $F_{scan}\in R^{c\times h \times w}.$8. Adding 2DSSA changes these to IoU $F_{scan}\in R^{c\times h \times w}.$9, CD $$F_{down}\in R^{2c\times \frac{h}{2} \times \frac{w}{2}},$$0, and MAE $$F_{down}\in R^{2c\times \frac{h}{2} \times \frac{w}{2}},$$1, corresponding to gains of $$F_{down}\in R^{2c\times \frac{h}{2} \times \frac{w}{2}},$$2 in IoU, $$F_{down}\in R^{2c\times \frac{h}{2} \times \frac{w}{2}},$$3 in CD, and $$F_{down}\in R^{2c\times \frac{h}{2} \times \frac{w}{2}},$$4 in MAE (Chen et al., 23 Jul 2025). On nuScenes, the baseline reports IoU $$F_{down}\in R^{2c\times \frac{h}{2} \times \frac{w}{2}},$$5, CD $$F_{down}\in R^{2c\times \frac{h}{2} \times \frac{w}{2}},$$6, and MAE $$F_{down}\in R^{2c\times \frac{h}{2} \times \frac{w}{2}},$$7, while baseline plus 2DSSA reports IoU $$F_{down}\in R^{2c\times \frac{h}{2} \times \frac{w}{2}},$$8, CD $$F_{down}\in R^{2c\times \frac{h}{2} \times \frac{w}{2}},$$9, and MAE $F_{up}\in R^{\frac{c}{2}\times 2h \times 2w}.$0, corresponding to $F_{up}\in R^{\frac{c}{2}\times 2h \times 2w}.$1 in IoU, $F_{up}\in R^{\frac{c}{2}\times 2h \times 2w}.$2 in CD, and a slight MAE worsening of $F_{up}\in R^{\frac{c}{2}\times 2h \times 2w}.$3 (Chen et al., 23 Jul 2025).

These ablations are notable because the largest improvements appear in Chamfer Distance. This suggests that 2DSSA contributes more strongly to geometric fidelity than to raw range-image reconstruction error. The paper summarizes the effect by stating that “By applying 2DSSA to the training, the model dynamically focuses on informative spatial regions, enhancing its ability to capture fine-grained geometric structures” (Chen et al., 23 Jul 2025).

For the full SRMambaV2 system, KITTI-360 results are IoU 0.4516, CD 0.0826, MAE 0.0041, and JSD 0.0045. The corresponding SRMamba baseline is IoU $F_{up}\in R^{\frac{c}{2}\times 2h \times 2w}.$4, CD $F_{up}\in R^{\frac{c}{2}\times 2h \times 2w}.$5, MAE $F_{up}\in R^{\frac{c}{2}\times 2h \times 2w}.$6, and JSD $F_{up}\in R^{\frac{c}{2}\times 2h \times 2w}.$7. On nuScenes, SRMambaV2 reports IoU 0.3299, CD 0.9485, MAE 0.0284, and JSD 0.0267, compared with SRMamba at IoU $F_{up}\in R^{\frac{c}{2}\times 2h \times 2w}.$8, CD $F_{up}\in R^{\frac{c}{2}\times 2h \times 2w}.$9, MAE $I_{latent}$0, and JSD $I_{latent}$1 (Chen et al., 23 Jul 2025). These are whole-system results rather than pure 2DSSA measurements, but the paper attributes a meaningful portion of the gains to improved handling of sparse distant structure.

The qualitative evidence is aligned with the same interpretation. In sparse nuScenes scenes, the paper reports that under only two laser beams on a car, SMFA and TULIP reconstruct only part of the front region, SRMamba reconstructs the basic contour but adds spurious noise, and SRMambaV2 reconstructs the complete contour while suppressing artifact noise (Chen et al., 23 Jul 2025). The depth-stratified analysis is also described as showing consistent outperformance across depth intervals, with particularly strong CD improvement.

Taken together, the reported evidence characterizes 2DSSA as a module whose primary observed contribution is improved structural reconstruction in distant sparse regions, especially through better contour recovery and reduced spurious noise.

6. Position within the broader 2D selective-scanning literature

The broader relevance of 2DSSA becomes clearer when placed against adjacent work on vision Mamba and selective visual computation. Analysis of vision-based Mamba models shows that when 2D images are serialized into 1D sequences, the learned inter-patch interaction pattern is strongly shaped by scan order: models tend to attend to preceding patches, with “preceding” defined by the chosen traversal path (Wang et al., 28 Feb 2025). This matters because 2DSSA explicitly uses four directional scans rather than a single serialization, which can be read as an attempt to reduce the anisotropy induced by any one scan route.

Related work on multi-route visual scanning reaches a similar conclusion from a different angle. MHS-VM argues that 2D image understanding can be approximated by combining several complementary sequential views of the same feature map, and states that “These different 1D causal relationships between patches are combined to approximate the complex structural relationships in 2D vision data” (Ji, 2024). Although MHS-VM is not a self-attention method, its emphasis on route diversity provides context for understanding why 2DSSA employs multi-directional SS2D before Swin-based focusing.

By contrast, ToSA is selective over 2D image tokens but does not implement scanning. It selects subsets of patch tokens for self-attention and lets the remaining tokens bypass the layer, preserving the full token set for dense prediction, but it does not define a scan order, traversal dynamics, or recurrent scan state (Singh et al., 2024). Relative to this design space, 2DSSA occupies a distinct position: it combines scan-based global browsing with subsequent attention, rather than performing token subset selection alone.

The main limitations of 2DSSA, as documented, are specification-level rather than conceptual. The SRMambaV2 paper does not provide the internal SS2D equations, the state-space recurrence, the selective gate formulation, the directional fusion rule, the exact tensor layout entering Swin after modulation, an explicit sparse mask over valid LiDAR pixels, or module-level FLOPs, memory, and latency (Chen et al., 23 Jul 2025). It also offers a conceptual efficiency rationale—SS2D for global modeling and Swin instead of full ViT because quadratic global attention is impractical for large super-resolution images—but does not report a dedicated efficiency analysis for 2DSSA itself.

In that sense, Biomimetic 2D Selective Scanning Self-Attention is best understood as a hybrid module family member rather than as a fully formalized standalone operator. Its distinctive contribution lies in coupling four-direction selective scanning, saliency-oriented modulation, and context-guided attention under a biomimetic coarse-to-fine rationale tailored to the depth-structured sparsity of LiDAR range images.