Motion-Guided Spatial Sparsification

• Motion-guided spatial sparsification is a methodology that uses motion cues, such as optical flow and pose-derived signals, to allocate spatial capacities dynamically.
• Techniques range from hard pruning with binary masks to soft attention reweighting, optimizing resource usage in dynamic regions while preserving contextual information.
• Empirical studies show these methods enhance computational efficiency and accuracy in tasks like crowd counting, video generation, and SLAM by adaptively focusing on motion-relevant areas.

Motion-guided spatial sparsification denotes a class of methods in which motion cues determine where a model should allocate spatial capacity, whether by pruning tokens, restricting attention neighborhoods, selecting memory, modulating features, or preserving density only in dynamic regions. In the literature, the guiding signal may be optical flow, people flow, pose-derived motion flow, event activity, temporal feature differences, or pose-visibility geometry, and the resulting sparsity may be hard, as in binary token masks or frame selection, or soft, as in motion-biased warping and attention reweighting (Kandala et al., 2024, Bai et al., 2021, Fang et al., 2023, Yang et al., 4 Sep 2025, Yang et al., 30 Apr 2025, Park et al., 2022, Wuwu et al., 2 Jun 2026).

1. Conceptual scope and taxonomy

A central distinction in the literature is between hard sparsification and soft sparsification. Hard variants explicitly discard regions, tokens, or primitives. FocusMamba constructs binary sparsification maps $M_I$ and $M_E$ for RGB and event tokens through adaptive thresholding, so that low-information regions are discarded before backbone processing and fusion (Yang et al., 4 Sep 2025). MoSAM performs temporal sparsification by selecting only reliable past frames for memory and spatial sparsification by retaining only high-confidence foreground pixels in those frames (Yang et al., 30 Apr 2025). Dynamic Spatial Sparsification for vision backbones defines a binary decision mask $\hat{\mathbf{D}}$ and progressively prunes tokens in ViTs or routes locations to fast and slow paths in CNNs and hierarchical ViTs (Rao et al., 2022). SparseStreet learns binary Gaussian masks $M_k \in \{0,1\}$ and later prunes only background Gaussians according to global importance, leaving dynamic-node Gaussians untouched in the second stage (Wuwu et al., 2 Jun 2026). In visual SLAM, point sparsification is posed as a minimum-cost maximum-flow problem, and only points whose source-to-point flow exceeds a threshold are retained (Park et al., 2022).

Soft variants, by contrast, preserve dense representations but bias them toward motion-relevant support. Monet “does not explicitly prune computation,” yet it uses people flow to coarsely segment regions where a person may be and then reweights features through residual attention blocks, which the paper describes as a form of motion-guided spatial focusing (Bai et al., 2021). Pix2Gif does not use the term “sparsification,” but its motion-guided latent flow field, warped latent $z_W$, and perceptual regularization encourage localized changes while static regions remain close to the source (Kandala et al., 2024). MMN similarly introduces motion-guided modulation factors $\gamma_s,\beta_s,\gamma_t,\beta_t$ that act as soft spatial-temporal gates over skeleton features rather than as hard joint selection (Gu et al., 29 Jul 2025). DMGAL uses motion-derived self- and cross-association matrices whose softmax weights emphasize motion-related patches within videos and across task episodes, again yielding soft spatial sparsification rather than explicit pruning (Guo et al., 2024).

This division clarifies a common misconception: motion-guided spatial sparsification is not synonymous with hard token dropping. In published systems it includes hard pruning, masked memory, block-sparse attention, feature warping, affine modulation, and graph-structured capacity allocation, provided that motion determines where computation or representational emphasis is concentrated.

2. Motion signals and sparsification operators

The motion signal itself varies substantially across tasks. In crowd counting, Monet computes optical flow with PWCNet between two consecutive frames,

$$\mathbf{f}_T = \text{PWCNet}(I_{T-1}, I_T) \in \mathbb{R}^{H \times W \times 2},$$

and a segmentation network predicts a coarse people mask $M_T = S(\mathbf{f}_T; \theta_{\text{seg}})$, supervised by a pixel-wise BCE term and coupled to density regression through

$$L_{\text{total}} = L_{\text{den}} + \lambda L_{\text{seg}}.$$

The mask acts as a spatial prior that emphasizes likely people regions and suppresses background before density estimation (Bai et al., 2021).

In image-to-GIF generation, Pix2Gif derives a scalar motion magnitude from optical flow and learns a separate motion embedding $c_M$, combined with text as $M_E$0. The core spatial operator is a latent-space flow predictor

$M_E$1

followed by warped latent generation

$M_E$2

The perceptual regularizer

$M_E$3

constrains $M_E$4 to remain in the same semantic content space as the source, which, in the paper’s interpretation, encourages minimal, motion-focused edits (Kandala et al., 2024).

In dance video generation, Dance Your Latents uses RAFT on DensePose images to produce dense motion flow $M_E$5 and aligns tokens to a reference frame within each temporal subspace: $M_E$6 Subspace attention is then computed only within motion-aligned local cubes rather than over the full spatiotemporal volume, making the sparse attention pattern follow body-part trajectories rather than static grid neighborhoods (Fang et al., 2023).

In RGB-event detection, FocusMamba derives a motion-conditioned global sparsity control from the event stream: $M_E$7 This event spatial ratio modulates both score contrast and thresholding: $M_E$8 after which binary masks are formed by comparing token scores to $M_E$9 (Yang et al., 4 Sep 2025).

In video object segmentation, MoSAM uses both sparse and dense motion representations. Sparse motion is encoded by keypoints extracted from features and masks; dense motion is encoded by optical flow $\hat{\mathbf{D}}$0, masked by the current object region,

$\hat{\mathbf{D}}$1

and used to warp the current mask into the next frame, producing a motion-derived box prompt (Yang et al., 30 Apr 2025).

In SLAM and dynamic scene reconstruction, the “motion” signal is not image flow but camera and object dynamics. The SLAM sparsification method uses pose visibility, frame baselines, and spatial diversity; for a frame pair it defines

$\hat{\mathbf{D}}$2

and uses this baseline in the edge cost of a minimum-cost maximum-flow graph (Park et al., 2022). SparseStreet instead uses scene-graph node types and temporal variability: dynamic actors receive weak sparsity penalties, the background receives strong sparsity pressure, and time-dependent masks

$\hat{\mathbf{D}}$3

allow a Gaussian to be active only when needed (Wuwu et al., 2 Jun 2026).

These formulations show that “motion guidance” need not mean optical flow specifically. It can denote any signal that encodes change, visibility across poses, or dynamic object identity strongly enough to determine spatial allocation.

3. Architectural patterns

Several recurrent architectural patterns emerge. One is motion-guided masking before high-capacity reasoning. Monet first estimates coarse people regions from people flow and only then applies non-local spatial-temporal reasoning and residual attention refinement (Bai et al., 2021). FocusMamba follows the same sequence at token level: Event-Guided Multimodal Sparsification produces $\hat{\mathbf{D}}$4 and $\hat{\mathbf{D}}$5, and only the retained regions are processed by sparse VSS blocks and Cross-Modality Focus Fusion (Yang et al., 4 Sep 2025). MoSAM likewise performs motion-guided prompting and then restricts memory to selected frames and selected pixels before cross-attention is applied in SAM2 (Yang et al., 30 Apr 2025).

A second pattern is motion-guided local neighborhood restriction. Dance Your Latents decomposes the latent volume into non-overlapping regular 3D subspaces, applies attention only within each subspace,

$\hat{\mathbf{D}}$6

and then uses motion-flow-guided Align and Restore to convert irregular motion trajectories into regular attention neighborhoods in the aligned space (Fang et al., 2023). A plausible implication is that block-sparse attention becomes more effective when the sparse graph follows motion correspondences rather than a static spatial partition.

A third pattern is motion as a control signal rather than a direct input stream. MMN computes feature-space differences

$\hat{\mathbf{D}}$7

then transforms them into FiLM-like modulation parameters via MSM and MTM, using

$\hat{\mathbf{D}}$8

The paper explicitly interprets this as soft spatial sparsification because joints with negligible motion receive nearly neutral modulation (Gu et al., 29 Jul 2025). DMGAL uses the same control principle at patch level: motion features extracted from frame differences define self- and cross-association matrices that govern which patches aggregate from which other patches within a video or across a few-shot task (Guo et al., 2024).

A fourth pattern is progressive sparsification with cumulative decisions. Dynamic Spatial Sparsification for ViTs estimates token importance at several layers, updates a cumulative mask

$\hat{\mathbf{D}}$9

and uses masked attention during training and actual token removal during inference (Rao et al., 2022). The same paper generalizes the idea to CNNs and hierarchical ViTs by preserving the feature map grid and routing different locations through fast and slow paths. This framework is explicitly source-agnostic with respect to importance estimation, which suggests that motion cues can be injected without changing the basic sparsification machinery.

A fifth pattern is capacity preservation in dynamic regions. SparseStreet uses node-aware regularization

$M_k \in \{0,1\}$0

with stronger sparsity pressure on the background than on rigid, deformable, or SMPL nodes, and then performs background-only pruning in its second stage (Wuwu et al., 2 Jun 2026). This implements the inverse of typical image sparsification: the representation remains dense where motion and temporal consistency demand it, and becomes sparse where temporal redundancy is high.

4. Task domains and representative uses

In controllable generation, motion-guided sparsification appears in multiple forms. Guided Motion Diffusion introduces spatial constraints into human motion synthesis through a feature projection scheme, an imputation formulation, and a dense guidance approach that turns sparse signals such as sparse keyframes into denser guidance signals for reverse diffusion (Karunratanakul et al., 2023). Pix2Gif uses motion magnitude prompts to drive latent warping, thereby localizing image-to-video changes to motion-relevant regions (Kandala et al., 2024). Dance Your Latents uses motion-flow-guided subspace alignment to keep local attention coherent along body-part motion trajectories in dance videos (Fang et al., 2023). Diverse Human Motion Prediction Guided by Multi-Level Spatial-Temporal Anchors factorizes anchors into spatial anchors and temporal anchors so that different motion modes can selectively emphasize different joints and frequency patterns within an interaction-enhanced STGCN (Xu et al., 2023).

In video analysis, Monet applies motion-guided segmentation and non-local reasoning to crowd counting, while MMN and DMGAL use motion-derived modulation or attention to focus recognition on subtle but discriminative joints or patch regions (Bai et al., 2021, Gu et al., 29 Jul 2025, Guo et al., 2024). Monet’s people flow map is explicitly described as a prior on where people may exist; MMN’s visualization shows that feature maps increasingly highlight a few key joints; DMGAL’s S-MGA and C-MGA identify and correlate motion-related region features at both video and task levels.

In interactive segmentation and memory-based tracking, MoSAM uses motion-guided point and box prompts to bias the decoder toward the predicted future object region and uses Spatial-Temporal Memory Selection to keep only reliable frames and pixels in memory (Yang et al., 30 Apr 2025). Here sparsification is not merely computational; it is also epistemic, since unreliable memory entries are removed to reduce error accumulation.

In multimodal detection, FocusMamba uses the event modality to guide adaptive collaborative sparsification of both RGB and event tokens before fusion, then uses logical relations among the sparsification masks to identify complementary regions during cross-modal fusion (Yang et al., 4 Sep 2025). The method is explicitly motivated by the observation that background in images and non-event regions in event data should not be processed uniformly.

In 3D dynamic scene representations, SparseStreet uses scene-graph dynamics to preserve dense Gaussian coverage on vehicles, pedestrians, and other dynamic nodes while aggressively compressing the static background (Wuwu et al., 2 Jun 2026). In SLAM, point sparsification uses pose visibility, long-baseline constraints, and spatial diversity to keep the subset of landmarks that best preserves the bundle-adjustment problem structure (Park et al., 2022).

Across these domains, motion-guided spatial sparsification is therefore not tied to a single data type or architecture. It spans latent diffusion, GCNs, video transformers, memory-augmented segmenters, state-space multimodal detectors, Gaussian splatting, and graph-optimized SLAM.

5. Empirical evidence and recurring claims

Published results consistently associate motion-guided sparsification with better efficiency, better robustness, or both. In video crowd counting, Monet improves VidCrowd performance from MAE/MSE $M_k \in \{0,1\}$1 for the baseline and $M_k \in \{0,1\}$2 for the baseline with non-local blocks to $M_k \in \{0,1\}$3 for the full motion-guided model; on Mall it reaches $M_k \in \{0,1\}$4, and on UCSD $M_k \in \{0,1\}$5 (Bai et al., 2021). These results support the paper’s claim that motion-guided spatial focusing complements non-local spatial-temporal modeling rather than replacing it.

In dance generation, the progression from DisCo to video diffusion, then to STSA, then to motion-flow-guided Align&Restore yields FVD $M_k \in \{0,1\}$6, while FID-VID also improves (Fang et al., 2023). This sequence is especially instructive because the gains appear in lockstep with increasingly structured sparsity: first temporal coupling, then local subspace attention, then motion-aligned local attention.

In RGB-event detection, FocusMamba-B improves mAP from $M_k \in \{0,1\}$7 to $M_k \in \{0,1\}$8 while reducing FLOPs from $M_k \in \{0,1\}$9G to $z_W$0G; the EGMS component alone reduces FLOPs from $z_W$1G to $z_W$2G with mAP improvement of $z_W$3 (Yang et al., 4 Sep 2025). The paper also reports that fixed kept-rate sparsification at the same average FLOPs is inferior to event-guided adaptive kept ratios, which directly supports the claim that motion activity should modulate sparsity budgets sample by sample.

In video object segmentation, MoSAM raises LVOS-v1 performance for SAM2-L from $z_W$4 to $z_W$5, with ablation showing monotonic gains from sparse motion, dense motion, temporal selection, and spatial selection (Yang et al., 30 Apr 2025). This indicates that motion guidance and memory sparsification are complementary rather than competing mechanisms.

In skeleton-based micro-action recognition, MMN reaches Body Top-1 $z_W$6, Action Top-1 $z_W$7, Action Top-5 $z_W$8, and $z_W$9 on MA-52 in the 2S setting, while the full MSM+MTM design improves $\gamma_s,\beta_s,\gamma_t,\beta_t$0 from $\gamma_s,\beta_s,\gamma_t,\beta_t$1 to $\gamma_s,\beta_s,\gamma_t,\beta_t$2 in joint-only ablation (Gu et al., 29 Jul 2025). In few-shot action recognition, DMGAL-FT with TRX improves SSv2-Full 1-shot accuracy from $\gamma_s,\beta_s,\gamma_t,\beta_t$3 to $\gamma_s,\beta_s,\gamma_t,\beta_t$4, and its ablations show substantial gains from S-MGA and C-MGA individually and jointly (Guo et al., 2024).

In efficiency-centered backbone sparsification, DynamicViT reduces DeiT-S from $\gamma_s,\beta_s,\gamma_t,\beta_t$5 to $\gamma_s,\beta_s,\gamma_t,\beta_t$6 GFLOPs with top-1 accuracy changing from $\gamma_s,\beta_s,\gamma_t,\beta_t$7 to $\gamma_s,\beta_s,\gamma_t,\beta_t$8, and reports throughput gain of $\gamma_s,\beta_s,\gamma_t,\beta_t$9 (Rao et al., 2022). In SLAM, the abstract reports more accurate camera poses with approximately $$\mathbf{f}_T = \text{PWCNet}(I_{T-1}, I_T) \in \mathbb{R}^{H \times W \times 2},$$0 of the map points and $$\mathbf{f}_T = \text{PWCNet}(I_{T-1}, I_T) \in \mathbb{R}^{H \times W \times 2},$$1 of the computation (Park et al., 2022). SparseStreet reports up to $$\mathbf{f}_T = \text{PWCNet}(I_{T-1}, I_T) \in \mathbb{R}^{H \times W \times 2},$$2 compression ratio with minimal quality degradation and, in the Waymo setting built on OmniRe, roughly $$\mathbf{f}_T = \text{PWCNet}(I_{T-1}, I_T) \in \mathbb{R}^{H \times W \times 2},$$3M to $$\mathbf{f}_T = \text{PWCNet}(I_{T-1}, I_T) \in \mathbb{R}^{H \times W \times 2},$$4M Gaussians with $$\mathbf{f}_T = \text{PWCNet}(I_{T-1}, I_T) \in \mathbb{R}^{H \times W \times 2},$$5 to $$\mathbf{f}_T = \text{PWCNet}(I_{T-1}, I_T) \in \mathbb{R}^{H \times W \times 2},$$6 FPS (Wuwu et al., 2 Jun 2026).

A recurring misconception is that sparsification inevitably trades away fidelity. The literature is more specific: under moderate or structured sparsification, several systems report improved accuracy or consistency because motion-guided pruning removes background interference, poor memory, or redundant static capacity rather than deleting useful evidence.

6. Limitations, assumptions, and open directions

The major limitations are task-dependent but structurally similar. FocusMamba explicitly notes that when the scene and the object remain stationary relative to the camera, the event spatial ratio $$\mathbf{f}_T = \text{PWCNet}(I_{T-1}, I_T) \in \mathbb{R}^{H \times W \times 2},$$7 is difficult to estimate the object information; in that case the control factor approaches $$\mathbf{f}_T = \text{PWCNet}(I_{T-1}, I_T) \in \mathbb{R}^{H \times W \times 2},$$8 and loses its regulatory effect on token selection (Yang et al., 4 Sep 2025). Dance Your Latents relies on accurate pose estimation and RAFT flow on DensePose, uses fixed subspace sizes, and applies nearest-neighbor discretization $$\mathbf{f}_T = \text{PWCNet}(I_{T-1}, I_T) \in \mathbb{R}^{H \times W \times 2},$$9, so alignment quality is limited by motion-estimation accuracy and quantization (Fang et al., 2023). SparseStreet relies on a supervised scene graph and has not yet been integrated with self-supervised dynamic/static decomposition (Wuwu et al., 2 Jun 2026). MMN explicitly uses only soft modulation and no hard sparsity constraint, while its motion consistency is realized by multi-scale aggregation rather than an explicit auxiliary loss (Gu et al., 29 Jul 2025). Dynamic Spatial Sparsification shows that overly aggressive keep ratios cause rapid accuracy degradation, especially when $M_T = S(\mathbf{f}_T; \theta_{\text{seg}})$0 in DynamicViT or when inappropriate stages are sparsified in ConvNeXt and Swin (Rao et al., 2022).

Several open directions recur across the surveyed work. One is to convert soft motion cues into explicit sparse masks. Pix2Gif’s discussion of thresholding a latent flow magnitude map into a motion mask is presented as a concrete extension toward explicit sparse computation in motionless regions (Kandala et al., 2024). Another is to use motion guidance for sparse attention rather than only token pruning, as already exemplified by motion-aligned subspace attention in dance generation and motion-related cross-attention in few-shot action recognition (Fang et al., 2023, Guo et al., 2024). A third is to make sparsity temporally adaptive: MoSAM’s frame-level selection, SparseStreet’s time-dependent Gaussian masks, and DynamicViT’s progressive decisions point toward systems that decide not only where but also when dense computation is necessary (Yang et al., 30 Apr 2025, Wuwu et al., 2 Jun 2026, Rao et al., 2022).

A plausible implication of this trajectory is that future systems will couple motion estimation, spatial selection, and downstream prediction more tightly, rather than treating motion as a side input to otherwise fixed dense architectures. The surveyed literature already shows the essential ingredients: motion-derived saliency, adaptive sparsity budgets, structure-preserving sparse operators, and regularizers that prevent dynamic regions from being over-pruned.