Sparse Submanifold CNNs

Updated 7 July 2026

Sparse Submanifold Convolutional Neural Networks (SSCNNs) are specialized CNNs designed to operate on high-dimensional data with few active sites by computing features only where needed.
They use a submanifold convolution operation that preserves the active set across stride-1 layers, preventing unwanted support dilation seen in standard sparse convolutions.
SSCNNs have been successfully applied to diverse domains like detector imaging, medical CT, and point cloud processing, achieving significant improvements in memory usage and processing speed.

Sparse Submanifold Convolutional Neural Networks (SSCNNs), also called Submanifold Sparse Convolutional Networks (SSCNs), are convolutional neural networks designed for inputs in which only a small fraction of pixels or voxels are active, while the active sites lie on thin, connected, lower-dimensional structures embedded in a higher-dimensional grid. Their defining operation is the submanifold sparse convolution, which computes features only at active sites and, in the stride-1 case, preserves the active set across layers rather than allowing activations to spread into empty space. This formulation was introduced by Graham and van der Maaten for sparse high-dimensional data such as pen strokes, point clouds, and sparse spatio-temporal signals, and was later developed into semantic-segmentation architectures and generalized sparse-tensor frameworks for 3D and 4D perception (Graham et al., 2017, Graham et al., 2017, Choy et al., 2019).

1. Origins and conceptual scope

The original motivation for SSCNNs is that many scientific and geometric data sources are intrinsically sparse. Examples stated in the literature include handwritten pen strokes in 2D, 3D point clouds from LiDAR or RGB-D sensors, surfaces in voxel grids, and sparse spatio-temporal data in 4D. Dense CNNs are inefficient in these regimes because they store and process every location in the ambient grid, even though most sites are in a ground state or zero state. The problem worsens with dimension, because the enclosing grid grows exponentially while the signal often remains confined to a lower-dimensional support (Graham et al., 2017, Graham et al., 2017).

A second motivation is architectural rather than representational. Earlier sparse convolutional methods already avoided explicit computation on inactive sites, but they typically determined output activity using the usual convolutional rule: if any input in the receptive field is active, the output site can become active. Repeated application therefore causes “submanifold dilation,” progressively densifying the feature maps. In the formulation highlighted by later LArTPC work, even a single nonzero site can expand to $3^d$ nonzero sites after one dense convolution and $5^d$ after two convolutions, where $d$ is the spatial dimension. SSCNNs were introduced specifically to avoid this support explosion while retaining deep convolutional processing (Graham et al., 2017, Dominé et al., 2019).

The terminology evolved across papers. The 2017 paper "Submanifold Sparse Convolutional Networks" defined Sparse Convolution (SC) and Valid Sparse Convolution (VSC), with VSC preserving the active set exactly. The later paper "3D Semantic Segmentation with Submanifold Sparse Convolutional Networks" adopted the now-standard terminology submanifold sparse convolution (SSC) for the sparsity-preserving operator. The Minkowski framework then embedded this construction into a generalized sparse convolution over arbitrary input and output coordinate sets, making the submanifold case one specific choice, namely $\mathcal{C}^{\text{out}} = \mathcal{C}^{\text{in}}$ (Graham et al., 2017, Graham et al., 2017, Choy et al., 2019).

2. Operators, active sets, and sparse-tensor mechanics

An SSCNN operates on a sparse tensor: a set of active coordinates together with feature vectors attached only to those coordinates. A site is active if its feature vector is not in the ground state, usually nonzero. This is the central departure from dense CNNs, where a full lattice tensor is materialized regardless of occupancy (Graham et al., 2017, Choy et al., 2019).

The core operators are summarized below.

Operator	Activity rule	Consequence
Sparse convolution $SC$	Output can be active if any active input lies in the receptive field	Saves computation on zeros, but can densify support
Submanifold sparse convolution $SSC$ / Valid Sparse Convolution $VSC$	Output is active if and only if the center site is active	Preserves the active set across stride-1 layers
Generalized sparse convolution	Input and output coordinate sets are explicit and need not match	Unifies dense, sparse, strided, transposed, and submanifold cases

In one notation used for LArTPC reconstruction, the two principal operators are $SC(n,m,f,s)$ and $SSC(n,m,f)$ , where $n$ is the number of input channels, $5^d$ 0 the number of output channels, $5^d$ 1 the filter size, and $5^d$ 2 the stride. Sparse convolution behaves like a standard convolution on the sparse representation, with inactive input pixels treated as zero; if the input size is $5^d$ 3, the output size is $5^d$ 4. Submanifold sparse convolution is the stride- $5^d$ 5 padded case in which the output size matches the input size and an output pixel is nonzero if and only if the center pixel of its receptive field is nonzero in the input feature map. The operational summary given in that work is $5^d$ 6 in the stride-1 case (Dominé et al., 2019).

The implementation described in the foundational semantic-segmentation work stores each sparse layer as a hash table mapping coordinates to rows in a feature matrix of size $5^d$ 7, where $5^d$ 8 is the number of active sites and $5^d$ 9 the number of input feature planes. Convolution is driven by a rule book $d$ 0, with $d$ 1, where each $d$ 2 contains input-output index pairs for a given kernel offset. Minkowski Engine expresses the same idea through kernel maps, index lists that gather active inputs, apply offset-specific weight matrices, and scatter-accumulate into active outputs. In both formulations, the sparse pattern is encoded by coordinates and index maps rather than by a dense mask (Graham et al., 2017, Choy et al., 2019).

The same sparsity discipline is extended to non-convolutional components. In LArTPC SSCNNs, activation functions and batch normalization are redefined by restricting them to the nonzero pixels only. This is important because the computational advantage of SSCNNs depends not only on sparse convolution but on maintaining sparse intermediate representations throughout the network (Dominé et al., 2019).

A frequent misconception is that any sparse convolution is already “submanifold.” The literature distinguishes them sharply. Sparse convolution reduces cost relative to dense convolution, but it does not by itself solve support dilation. The submanifold restriction is the mechanism that prevents the active set from filling neighboring empty regions layer after layer (Graham et al., 2017, Graham et al., 2017).

3. Architectural patterns and software realizations

Because SSC layers preserve sparsity too strictly to communicate across disconnected components, practical SSCNNs combine two regimes. Within a resolution stage, repeated SSC or VSC layers preserve the current active manifold. Between stages, strided sparse convolutions, pooling, or related transitions alter the support, enlarge the receptive field, and allow separate structures to merge at coarser scales. The 2017 semantic-segmentation paper states this explicitly as a limitation of SSC alone: it cannot propagate information between disconnected components without pooling or strided convolutions (Graham et al., 2017).

This pattern supports sparse versions of standard CNN families. The early work demonstrated VGG-, ResNet-, and DenseNet-style networks for handwriting and 3D shape classification, while the semantic-segmentation paper built sparse FCN and U-Net architectures with pre-activated $d$ 3 layers, batch normalization, ReLU, and residual blocks. Downsampling used $d$ 4, and reconstruction used deconvolution or nearest-neighbor upsampling. These results established that deep residual and encoder-decoder architectures can be transferred to sparse settings without abandoning convolutional locality (Graham et al., 2017, Graham et al., 2017).

Later domain-specific systems retained the same architectural logic while changing the backbone. LArTPC segmentation used both dense and sparse U-ResNet variants, combining U-Net encoder-decoder structure with ResNet-style residual connections. CT ROI detection used a submanifold sparse U-Net-34 in Minkowski Engine v0.5.4. The KiTS23 kidney-tumour system used a custom sparse voxel-based 3D U-Net with ConvNeXtV2 blocks, MinkowskiConvolution, MinkowskiConvolutionTranspose, MinkowskiLayerNorm, skip connections by element-wise summation, and 27,473,696 parameters. Digital sun-sensor calibration used a modified ResNet-34 implemented with SpConv, in which sparse convolutions performed downsampling and SubMConv2d residual layers preserved support (Dominé et al., 2019, Alonso-Monsalve et al., 2022, Alonso-Monsalve et al., 6 Nov 2025, Herman et al., 29 Jul 2025).

Software ecosystems were correspondingly important. The LArTPC and MicroBooNE systems used SparseConvNet in PyTorch; several medical-imaging and neutrino-telescope systems used Minkowski Engine; DSS predictive calibration used SpConv. Minkowski Engine generalized sparse convolutions to arbitrary dimension and kernel support, including strided, dilated, transposed, and submanifold cases, and introduced hybrid kernels for 4D space-time processing (Dominé et al., 2019, collaboration et al., 2020, Choy et al., 2019).

4. Detector imaging and event reconstruction

Large particle-imaging detectors provided one of the clearest demonstrations of SSCNN scalability. In liquid argon time projection chamber data, the 2019 study reports that on average only about 0.01% of pixels in an event are non-zero, while the nonzero regions form long, narrow, continuous trajectories. On a 3D semantic-segmentation task, sparse U-ResNet reduced inference memory by a factor of 364 and wall-time by a factor of 33 relative to a standard CNN, without loss of accuracy; the corresponding factors for 2D samples were 93 and 3.1. For the 3D 192 px case, dense nonzero accuracy was 92% and sparse nonzero accuracy was 94%, and for larger sparse models overall accuracy reached roughly 99%. The same study also introduced a machine-learning pipeline for Michel-electron reconstruction: sparse U-ResNet segmentation, selection of predicted Michel-electron and MIP pixels, DBSCAN clustering, candidate selection by attachment to a predicted MIP cluster, and matching to true Michel clusters by maximum overlap. The reported Michel-electron identification efficiency was 93.9% with 96.7% true positive rate, and reconstructed clusters achieved 95.4% average pixel clustering efficiency and 95.5% purity (Dominé et al., 2019).

MicroBooNE translated the same paradigm into production semantic segmentation. SparseSSNet processed zero-suppressed wire-plane images of size $d$ 5, with only about $d$ 6 of pixels passing the threshold used to build the sparse input. It used three independently trained sparse U-Res-Nets, one per wire plane, to classify nonzero pixels into HIP, MIP, Shower, Delta rays, and Michel electrons, later regrouped into the analysis categories Track and Shower. The paper reports test-sample accuracy $d$ 7; for the collection plane, overall accuracy on nonzero pixels was about 96%, with 0.992 Track accuracy and 0.996 Shower accuracy. Relative to the previous dense CNN, the sparse model reduced processing time for a $d$ 8 image from about 5 s to about 0.5 s and memory usage from about 6 GB to about 1 GB (collaboration et al., 2020).

Neutrino telescopes extended SSCNNs to sparse, irregular, effectively 4D event data. In an IceCube-like detector, the input was represented as coordinates $d$ 9 of size $\mathcal{C}^{\text{out}} = \mathcal{C}^{\text{in}}$ 0 with $\mathcal{C}^{\text{out}} = \mathcal{C}^{\text{in}}$ 1 entries and features $\mathcal{C}^{\text{out}} = \mathcal{C}^{\text{in}}$ 2 containing photon-hit counts in a 1 ns window. The trigger-level sample had spatial sparsity about 2%, and quality-selected events about 3%. A ResNet-based SSCNN implemented in PyTorch and MinkowskiEngine used sparse submanifold convolutions, batch normalization, PReLU, and stride-2 sparse submanifold convolutions for downsampling. The abstract reports that the SSCNN ran approximately 16 times faster than a traditional CNN on a GPU. On the A100 GPU, the angular model reconstructed at 9901 Hz with batch size 12,288 events simultaneously, and achieved under 4° median angular resolution on the highest-energy trigger-level events; with quality cuts, performance improved by about 2° across the energy range (Yu et al., 2023).

5. Medical imaging, volumetric segmentation, and other application domains

In medical CT segmentation, SSCNNs were presented as an alternative to the standard practice of aggressive downsampling or patch-wise processing. A 2022 renal-cancer study sparsified CT volumes by Hounsfield-unit thresholding and then applied a two-stage SSCN pipeline: sparse input, coarse ROI detection, cropping of the original volume, and full-resolution sparse segmentation. For the final method, the threshold range was $\mathcal{C}^{\text{out}} = \mathcal{C}^{\text{in}}$ 3 H.U., with 2.08% signal loss and 76.63% background removal, reducing the number of voxels by about a factor of 2 per CT scan without downsampling. The first-stage submanifold sparse U-Net-34 achieved 84.6% Dice similarity coefficient, 99.1% accuracy, and training time of 2–3 minutes per epoch (Alonso-Monsalve et al., 2022).

The 2025 KiTS23 follow-up made the two-stage design fully explicit at high resolution. Stage 1 used CT scans resampled to 1.99 mm isotropic resolution, voxel sparsification by Hounsfield-unit thresholds derived from the 0.5th and 99.5th percentiles of kidney/mass voxels, a sparse 3D U-Net ROI detector, a probability threshold of 0.1, and elliptical dilation of size 11 voxels. Stage 2 used scans resampled to 0.78 mm isotropic resolution, ROI cropping, removal of disconnected predicted regions under 50 voxels, and separate processing of connected components. The sparse model achieved Dice scores of 95.8% for kidneys + masses, 85.7% for tumour + cyst, and 80.3% for tumour only, while reducing inference time by up to 60% and VRAM usage by up to 75% relative to an equivalent dense architecture (Alonso-Monsalve et al., 6 Nov 2025).

SSCNNs have also been applied to highly sparse optical sensing. In digital sun-sensor predictive calibration, a modified sparse ResNet-34 mapped a sparse $\mathcal{C}^{\text{out}} = \mathcal{C}^{\text{in}}$ 4 illumination image and a sub-field-of-view class label directly to the two sun angles $\mathcal{C}^{\text{out}} = \mathcal{C}^{\text{in}}$ 5. Sparse convolutions were used for downsampling and sparse submanifold convolutions for residual blocks, because DSS images contain only a small illuminated pattern on an otherwise zero background. On synthetic data, the system achieved a mean accuracy of 0.005° for the two sun-angle estimates; under mixed Poisson-Gaussian noise plus thresholding, the augmented model achieved mean error $\mathcal{C}^{\text{out}} = \mathcal{C}^{\text{in}}$ 6 with spread $\mathcal{C}^{\text{out}} = \mathcal{C}^{\text{in}}$ 7 for both single-aperture and multi-aperture configurations (Herman et al., 29 Jul 2025).

These applications show that SSCNNs are not confined to point clouds or detector voxelizations. The recurring pattern is sparsification or naturally sparse acquisition, followed by a sparse encoder-decoder or residual backbone that preserves support within stages and changes it only where the architecture requires receptive-field growth.

6. Efficiency claims, extensions, and limitations

Across the literature, the principal efficiency claim is consistent: SSCNN cost scales with the number of active sites rather than the full grid size. This is stated directly for LArTPC images, where larger detectors do not automatically imply prohibitive computation because sparse operations remain tied to the number of nonzero pixels. The same argument appears in MicroBooNE, where storage and convolution scale linearly with the number of selected pixels, in neutrino telescopes, where cost depends mainly on the number of OM hits, and in CT segmentation, where sparse tensors store only coordinates and features of active voxels (Dominé et al., 2019, collaboration et al., 2020, Yu et al., 2023, Alonso-Monsalve et al., 6 Nov 2025).

This does not mean that all redundancy disappears once submanifold convolution is introduced. The 2022 object-detection paper "Spatial Pruned Sparse Convolution for Efficient 3D Object Detection" argues that sparse 3D CNNs still waste substantial compute on background-dominated active voxels, and that regular sparse convolution in downsampling layers can amplify sparsity into denser intermediate representations. Its proposed SPS-Conv family—SPSS-Conv for submanifold blocks and SPRS-Conv for regular sparse downsampling blocks—uses magnitude-guided pruning to suppress redundant computation. On KITTI, Waymo, and nuScenes, the reported reduction is more than 50% reduction in GFLOPs without compromising the performance, with the Waymo example decreasing from 76.7 GFLOPs to 28.8 GFLOPs when both modules are used (Liu et al., 2022).

Several limitations recur. First, SSC alone cannot propagate information between disconnected components, so practical networks require strided sparse convolutions, pooling, or deconvolutional pathways. Second, the approach remains tied to voxelized or quantized sparse grids, so quantization and thresholding choices matter. Third, the advantage is strongest when sparsity is high; the 2017 semantic-segmentation paper states that dense inputs would not benefit as much. Fourth, some application papers emphasize domain-specific constraints rather than intrinsic limits of SSCNNs: the neutrino-telescope study did not include full electronics response inside the OM and used an approximate trigger logic; the DSS calibration system was trained on synthetic data and notes a Sim2Real gap; the KiTS23 study used only 489 training cases, had no access to official test labels, and reported final-epoch rather than best-validation results (Graham et al., 2017, Yu et al., 2023, Herman et al., 29 Jul 2025, Alonso-Monsalve et al., 6 Nov 2025).

A common misconception is therefore that SSCNNs are simply “sparse CNNs” in a generic sense. The literature treats the submanifold constraint as the decisive architectural feature: it prevents support growth, stabilizes memory and FLOP costs across deep stages, and makes large sparse scientific images tractable. Another misconception is that submanifold preservation alone is sufficient; in practice, SSCNNs are most effective as hybrid sparse architectures in which support-preserving blocks are interleaved with explicitly support-changing transitions. Within that design space, they have become a standard tool for high-dimensional sparse perception, spanning 2D pen strokes, 3D point clouds, 4D spatio-temporal data, detector imaging, medical CT, and sparse optical sensing (Graham et al., 2017, Graham et al., 2017, Choy et al., 2019).