Dynamic Resolution Multiscale Voxelization (DR-MSV)

• Dynamic Resolution Multiscale Voxelization (DR-MSV) is an adaptive strategy that varies voxel granularity based on local complexity to balance detail and efficiency.
• It employs metrics such as TSDF variance, point density, and reconstruction loss to selectively merge or refine voxel groups in different regions.
• Empirical results show that DR-MSV methods achieve significant speedups and memory savings while improving geometric reconstruction accuracy compared to fixed-resolution approaches.

Dynamic Resolution Multiscale Voxelization (DR-MSV) denotes an adaptive voxelization strategy in which voxel granularity is varied locally rather than fixed uniformly across a scene or point cloud. In the literature, the term appears explicitly in NeuroVoxel-LM as a “dynamic resolution multi-scale voxelisation” technique for large-scale point-cloud processing (Liu et al., 27 Jul 2025). Closely related mechanisms appear under different names in RGB-D supervised surface reconstruction through dynamic grid subdivision (Xu et al., 2023), in hash-based TSDF reconstruction through variance-adaptive multi-resolution voxel grids (Rebotti et al., 26 Nov 2025), and in point-cloud learning through on-the-fly dynamic voxelization with self-adaptive resolution (Su et al., 2020). Across these formulations, the shared objective is to preserve fine representation capacity in geometrically complex regions while coarsening homogeneous regions to improve computational and memory efficiency.

1. Definitional scope and representative formulations

DR-MSV, in its narrowest sense, refers to the NeuroVoxel-LM voxelization scheme that begins from an initial fine voxelization, evaluates voxel complexity, and then iteratively merges only non-complex $2 \times 2 \times 2$ blocks into coarser parent voxels (Liu et al., 27 Jul 2025). In a broader comparative sense, the same core principle is instantiated by systems that refine or merge voxels according to local TSDF variance, reconstruction loss, or local point density rather than by a single fixed scene-wide resolution (Rebotti et al., 26 Nov 2025, Xu et al., 2023, Su et al., 2020).

Formulation	Resolution-adaptation signal	Structural realization
NeuroVoxel-LM	Six geometric and structural indicators with 75th-percentile thresholds	Hierarchical voxel pyramid with iterative $2 \times 2 \times 2$ merging
Dynamic Grid Optimization	Surface-adjacent voxels with high image or SDF loss	Coarse grid with recursive subdivision into eight children
MrHash	Local TSDF variance	Multi-resolution voxel blocks stored in a flat spatial hash table
DV-ConvNet	Local density via k-NN and dilation	On-the-fly local $S^3$ voxel kernels per convolution layer

This suggests that DR-MSV is best understood not as a single data structure, but as a design principle for adaptive volumetric discretization. The principal design question is not whether voxels are used, but how local complexity is quantified, how resolution transitions are represented, and how lookup, update, and rendering or inference are made efficient.

2. Resolution adaptation criteria

In NeuroVoxel-LM, voxel complexity is evaluated from six indicators: point density $d$, surface roughness $\sigma_s$, normal coherence $\cos\theta$, structural principal component indicators $\lambda_{\text{linear}}$ and $\lambda_{\text{planar}}$, spatial distribution entropy $H_s$, and curvature $\kappa$ (Liu et al., 27 Jul 2025). The paper defines, for example,

$2 \times 2 \times 2$0

$2 \times 2 \times 2$1

$2 \times 2 \times 2$2

and

$2 \times 2 \times 2$3

Thresholds are not hand-tuned; a voxel is termed complex if its metric exceeds the 75th percentile of the related measure across the scene. Only when all eight children in a $2 \times 2 \times 2$4 block are non-complex are they merged into a coarser parent voxel.

In MrHash, the adaptive signal is statistical rather than explicitly geometric. Each voxel stores truncated signed distance $2 \times 2 \times 2$5, weight $2 \times 2 \times 2$6, color, and the variance of the mean TSDF, $2 \times 2 \times 2$7. The paper explicitly interprets low variance as indicating smooth, homogeneous regions and high variance as indicating boundaries, fine structure, noise, or complex geometry (Rebotti et al., 26 Nov 2025). TSDF observations are integrated by a running weighted average with $2 \times 2 \times 2$8, and variance is maintained with Welford’s online algorithm. The adaptive rule is variance-thresholded: high TSDF variance keeps voxel blocks at fine resolution, while blocks below a threshold are merged or downsampled to a coarser resolution.

In Dynamic Voxel Grid Optimization for RGB-D reconstruction, refinement is driven by optimization difficulty rather than local descriptive statistics. A voxel is a subdivision candidate only if its corner SDF values place it near the truncation or surface band, and it is then split when either the accumulated image loss or SDF loss exceeds $2 \times 2 \times 2$9 (Xu et al., 2023). This produces a loss-driven refinement policy concentrated on difficult surface-adjacent regions rather than on free space.

In DV-ConvNet, “dynamic resolution” has a different operational meaning. The network samples centroids by farthest point sampling, gathers $S^3$0 neighbors by k-NN, randomly picks $S^3$1 points, and fits them into a cubic local voxel kernel whose radius is chosen so that the farthest neighboring point can just be included (Su et al., 2020). Dense regions therefore induce smaller effective radii and finer local spatial focus, while sparse regions induce broader support. The method does not build a global adaptive scene grid; instead, it dynamically adapts local receptive fields at each convolution layer.

3. Structural realizations and data layout

NeuroVoxel-LM implements DR-MSV as an adaptive, hierarchical voxel pyramid. The pipeline begins with an initial fine voxelization such as 16 × 16 × 16, classifies voxels as complex or non-complex, and repeatedly merges only non-complex $S^3$2 groups (Liu et al., 27 Jul 2025). The result is a mixed-scale representation with small voxels in structurally complex regions and larger voxels in simple or homogeneous regions. The paper presents this as a direct alternative to fixed-resolution voxelization.

Dynamic Voxel Grid Optimization also uses an explicitly hierarchical representation, but its storage is organized around direct voxel-grid optimization rather than a separately named complexity classifier. The scene begins as a coarse grid

$S^3$3

and selected voxels are subdivided into eight children every $S^3$4 steps until $S^3$5 or until no further voxels qualify (Xu et al., 2023). Topology and modality storage are separated into $S^3$6, $S^3$7, and $S^3$8, where $S^3$9 stores node modalities, $d$0 stores pointers to eight corner nodes for each voxel, and $d$1 provides level-specific indexing. This separation is presented as a memory-efficient mechanism for multiresolution lookup and updates.

MrHash departs from hierarchical tree structures altogether. It avoids octrees and OpenVDB-style trees, which the paper characterizes as requiring recursive traversal, pointer-heavy layouts, $d$2 access or update costs, and poor GPU suitability (Rebotti et al., 26 Nov 2025). Instead, it stores all voxel blocks, regardless of resolution, in a single flat spatial hash table. Each hash entry stores block origin coordinates, a collision offset, a GPU pointer to the voxel block, and a resolution index indicating which resolution-specific GPU heap contains the block. The hash function is

$d$3

with $d$4, $d$5, and $d$6. The reported consequence is average constant-time access, a single address space for mixed resolutions, and full GPU parallelism.

DV-ConvNet uses neither a persistent global multiresolution grid nor a hash-based sparse map. Each DV-Conv layer voxelizes local neighborhoods into $d$7 cells, with $d$8 fixed in the reported system (Su et al., 2020). Features in occupied cells are aggregated by channel-wise max pooling, empty cells receive zero, and only the first five points encountered in a voxel are considered. This makes the representation sparse in a different sense: only local voxel kernels around sampled centroids are materialized.

4. Optimization, integration, and inference pipelines

In Dynamic Voxel Grid Optimization, DR-MSV is embedded in a direct optimization framework in which the scene stores SDF and radiance or color explicitly in the grid rather than in an MLP (Xu et al., 2023). For a query point $d$9, the paper interpolates

$\sigma_s$0

using trilinear interpolation. RGB-D supervision is given by

$\sigma_s$1

with $\sigma_s$2, $\sigma_s$3, and $\sigma_s$4. The reported training setup uses 60K iterations, subdivision at 20K and 40K, a learning rate of 0.1 with exponential decay by factor 0.1 at 20K, 2048 rays per iteration, and 128 uniformly sampled plus another 128 visibility-guided points per ray.

MrHash formulates adaptation around TSDF or SDF integration from either LiDAR or RGB-D observations. The pipeline executes allocation, integration, variance update, block downsampling or merging, and Marching Cubes surface extraction on the GPU (Rebotti et al., 26 Nov 2025). Allocation uses DDA traversal to identify intersected blocks, integration updates voxels in parallel, per-voxel variance is computed via parallel reductions, block variance is aggregated on the GPU, and resolution changes are applied without recursive traversal. The paper also extends the representation to GPU-accelerated rendering through a parallel quadtree for Gaussian Splatting, where local contrast in image tiles controls splat density.

DV-ConvNet integrates dynamic voxelization into a fully convolutional recognition pipeline rather than a reconstruction pipeline. Its forward pass consists of centroid sampling, k-NN neighborhood gathering, adaptive local voxelization, and 3D group convolution over the resulting $\sigma_s$5 local grids (Su et al., 2020). The voxelized tensor is

$\sigma_s$6

and the convolution kernel is

$\sigma_s$7

Indices of points contributing to each voxel are stored during forward propagation and reused during backward propagation, so the voxelization step is fully differentiable. Group convolution is defined over $\sigma_s$8 or $\sigma_s$9, extending rotational and reflectional symmetries to $\cos\theta$0.

Within NeuroVoxel-LM, DR-MSV is positioned as the point-cloud voxelization and geometric reconstruction component, complementary to NeRF-based representation and to the Token-level Adaptive Pooling for Lightweight Meta-Embedding mechanism (Liu et al., 27 Jul 2025). The paper explicitly states that DR-MSV handles point cloud feature extraction and geometric reconstruction accuracy, whereas TAP-LME processes NeRF weights for semantic representation. No DR-MSV-specific loss formula is provided in the text summarized here.

5. Empirical characteristics and trade-offs

The empirical literature describes DR-MSV-style methods primarily through the trade-off between efficiency and fidelity. In NeuroVoxel-LM, DR-MSV reduces total training time by over 35% relative to the fixed-resolution baseline, with TT decreasing from 19.87 h to 12.24 h, DP from 4.67 h to 2.56 h, MF from 15.20 s to 9.68 s, BT from 19.94 s to 12.28 s, and SP increasing from 0.81 shapes/s to 1.31 shapes/s (Liu et al., 27 Jul 2025). Reconstruction quality also improves: Chamfer Distance changes from 0.195 to 0.150, F1(PC) from 0.62 to 0.74, GeoIoU from 0.47 to 0.56, Acc from 0.69 to 0.74, Prec from 0.64 to 0.72, Rec from 0.61 to 0.71, F1(Vox) from 0.60 to 0.73, and VoxIoU from 0.45 to 0.56. Ablations show that DR-SV-MC and DR-MSV-BC are faster than full DR-MSV but less accurate, leading the authors to conclude that multiscale integration, multiclass supervision, and adaptive merging jointly provide the best balance.

MrHash reports up to 13× speedup and 4× lower memory usage compared to fixed-resolution baselines, while maintaining on-par reconstruction accuracy (Rebotti et al., 26 Nov 2025). For memory, the paper reports RGB-D savings from 2.0× to 7.5×, LiDAR savings from 2.3× to 2.9×, and average improvement from 1.95× to 4.07×. Runtime examples include 15.11 ms / 64.3 FPS on ScanNet for Ours (single) and 20.45 ms / 48.9 FPS on Replica. Reported average F-scores include 96.845 for RGB-D, 86.710 for LiDAR, and 93.804 overall on ScanNet + NC, and around 95.245 for RGB-D, around 83.131 for LiDAR, and around 89.188 overall on Replica + Oxford Spires. For Gaussian Splatting rendering, the reported multi-resolution result is PSNR 35.73, SSIM 0.960, LPIPS 0.044, and FPS 23.70, with a reconstruction mesh of 132.650 MB versus 454.025 MB for GSFusion on Replica.

Dynamic Voxel Grid Optimization reports that the dynamic grid can continue subdividing to level 4 while the fixed voxel grid reaches OOM at higher refinement levels (Xu et al., 2023). Relative to NeuralRGBD, the paper reports training reduced to roughly 50 minutes, and in a lighter setup about 20 minutes, while also reporting stronger geometric accuracy and improved recovery of thin objects, sharp edges, detailed surfaces, and complex scene geometry.

DV-ConvNet reports 146M FLOPs for the $\cos\theta$1 model and training time 0.054 s in the reported setup (Su et al., 2020). Its ablations indicate that max pooling inside voxels outperforms average pooling, k-NN adaptive sampling outperforms fixed-radius sampling, and $\cos\theta$2 with $\cos\theta$3 gives the best tested setting. The paper also notes several design limits: the kernel size is fixed to $\cos\theta$4, only the first five points in a voxel are considered, and on-the-fly voxel construction plus k-NN adds implementation overhead.

A recurrent trade-off across these papers is that more aggressive coarsening improves throughput and memory use but can degrade geometric fidelity in sparse or high-variance regions. MrHash makes this explicit: smaller variance thresholds preserve more fine voxels and better accuracy but use more memory, while larger thresholds reduce memory with slightly lower fidelity, especially under sparse LiDAR observations (Rebotti et al., 26 Nov 2025).

6. Relation to adjacent methods and common misunderstandings

A common misunderstanding is to equate DR-MSV with octree-based volumetric hierarchies. The comparative evidence does not support that restriction. MrHash is explicitly presented as a practical realization of the same adaptive idea without recursive octree traversal, replacing it with a flat spatial hash table and resolution-specific GPU heaps (Rebotti et al., 26 Nov 2025). This indicates that dynamic multiscale voxelization is compatible with non-hierarchical storage, provided that mixed resolutions can be indexed and updated efficiently.

A second misunderstanding is to treat DR-MSV as necessarily a global preprocessing stage. DV-ConvNet demonstrates a different interpretation in which voxelization occurs locally, on-the-fly, and at every convolution layer rather than once for the entire scene (Su et al., 2020). In that formulation, “dynamic resolution” refers to adaptive local support and per-layer receptive field control rather than to a persistent multiscale scene map.

A third misunderstanding is to conflate DR-MSV with language alignment in multimodal 3D models. In NeuroVoxel-LM, DR-MSV is the geometric efficiency and point-cloud feature extraction branch, whereas TAP-LME is the component that enhances semantic representation from NeRF weights (Liu et al., 27 Jul 2025). The paper does not present an explicit mathematical interface in which DR-MSV directly feeds NeRF losses or vice versa.

Finally, DR-MSV should not be understood as eliminating the fidelity–efficiency trade-off. The available evidence instead shows a controlled reallocation of resolution. In RGB-D reconstruction, the method refines high-loss surface-adjacent voxels and leaves easy regions coarse (Xu et al., 2023). In TSDF mapping, low-variance blocks are downsampled while high-variance blocks remain fine (Rebotti et al., 26 Nov 2025). In point-cloud learning, local neighborhoods are voxelized only where the network needs them (Su et al., 2020). The consistent technical theme is selective resolution allocation rather than universally higher resolution.

Taken together, these works place DR-MSV at the intersection of sparse volumetric representation, adaptive geometric modeling, and hardware-aware computation. Its main forms differ in the signals used for adaptation—complexity metrics, variance, loss, or density—and in the storage substrate—hierarchical grids, direct optimized voxel structures, flat hashes, or ephemeral local kernels. What remains invariant is the underlying principle: voxel resolution is treated as a local variable to be optimized where it counts.