{"id":"bfa50ca1-e7e6-42dd-bbec-035bed631e6b","arxiv_id":"2506.23157","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"STD-GS disentangles background and dynamic objects by clustering frame appearance and event motion features, and uses event brightness and flow to supervise Gaussian rendering, improving high-dynamic scene reconstruction.","lead":"This paper presents a Gaussian splatting method that uses both regular video frames and event cameras to reconstruct scenes with fast-moving objects. It separates static background from moving objects using clustering, then renders high-speed scenes more sharply than previous approaches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper provides no quantitative validation that the K-Means/ClusterFormer clusters correspond to actual background/object regions; if pseudo-labels are unreliable, the disentangled Gaussian decomposition and fusion are built on incorrect regions.","rationale":"I read the paper in good faith. The proposed pipeline is coherent, the ablations are internally consistent, and the use of events to compensate for frame limitations is plausible. The reader's verdict of CONDITIONAL is appropriate. However, the single most load-bearing assumption is the reliability of the clustering-based pseudo-labels. If the K-Means clusters do not correspond to actual background/object regions, then the Gaussian representation fusion, the EKF tracking, and the shadow-based fusion weights are all built on an incorrect decomposition, undermining the central claim that spatiotemporal disentanglement is what drives the performance gains. The paper provides only qualitative t-SNE plots and no validation of the segmentation quality. This is not an ad hominem or a disagreement with the authors' consensus; it is a missing support for a critical component. The concrete test I propose would directly settle whether the disentanglement is the source of the improvement or whether an oracle mask would change the result. I therefore keep the reader's conditional verdict unchanged, since the issue is addressable with additional experiments and reporting.","tokens_in":12546,"tokens_out":8974,"duration_ms":99714,"concrete_test":"Run the DSD module on HD-CED and Event-HyperNeRF with ground-truth or manually annotated foreground/background masks (annotatable for these datasets) and compute the IoU/accuracy of the K-Means/ClusterFormer pseudo-labels against these masks. Then train a variant of the full model where the pseudo-labels are replaced by the oracle masks, and compare PSNR/SSIM/LPIPS to the standard pipeline. If the oracle-mask variant improves significantly over the proposed pipeline, the clustering is a bottleneck; if it does not improve or degrades, the claimed benefit does not come from the disentanglement. Also report sensitivity to the number of clusters K (e.g., 2 vs. 3 or 4).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of time-continuous high-dynamic scene reconstruction depends on the Dynamic Scene Disentanglement (DSD) module correctly separating background from objects. DSD uses K-Means on SLIC superpixel appearance features and event correlation volumes to generate pseudo-labels, then trains ClusterFormer with the clustering loss in Eq. 1. However, no quantitative metric (IoU, accuracy) is reported against any ground-truth segmentation, and the t-SNE in Fig. 7 only shows that clustering separates abstract features, not that clusters correspond to background vs. object. Since K-Means always produces k separated clusters, the visualization is circular evidence. Second, Section 3.2 states 'we manually divide the regions of background and objects as candidate regions to match,' which is ambiguous. If manual masks are used to generate or refine the pseudo-labels, the method is not fully automatic and the disentanglement advantage may come from human annotations rather than the proposed appearance-motion clustering. Third, the number of clusters K is not specified; for scenes with multiple dynamic objects, a single K=2 clustering cannot disentangle multiple objects with different motions. If the masks are wrong, the 3D/4D Gaussian assignment (SfM points to background/object), the EKF tracking, and the shadow-based fusion in Eqs. 3-4 are built on incorrect regions, so the large gains over 4DGS (e.g., 27.12 vs. 20.60 PSNR on HD-CED) could arise from an implicit motion-segmentation prior rather than the claimed spatiotemporal disentanglement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes STD-GS, a multimodal Gaussian splatting framework for high-dynamic scene reconstruction that combines frame images and event streams. The method first disentangles the scene into static background and dynamic objects using a learnable appearance-motion clustering pipeline (Section 3.2), then represents the background as 3D Gaussians and dynamic objects as 4D Gaussians with EKF-based temporal tracking, and finally fuses the two sets through shadow-based weights (Section 3.3). The framework is trained with reconstruction, clustering, and spatiotemporal-consistency losses (Eqs. 1, 2, 5, 6). The authors evaluate on Event-HyperNeRF, HD-CED, and a newly captured CoFED dataset, reporting large gains over existing baselines, e.g., PSNR 27.12 versus 20.60 on HD-CED (Table 1).","tokens_in":12901,"tokens_out":4640,"duration_ms":55815,"significance":"If validated, the paper would make a useful contribution by showing that disentangling static and dynamic components in Gaussian splatting, with event data providing temporal cues, can improve reconstruction of high-dynamic scenes. The proposed CoFED dataset is a potentially valuable resource, and the ablations in Tables 3 to 6 are internally consistent with the contribution of each proposed loss and component. However, the central mechanism of the method—the reliability of the unsupervised background/object pseudo-labels—is not quantitatively verified, and parts of the reported superiority may be attributable to uncontrolled factors such as event-based deblurring. The paper also presents the event-Gaussian consistency as a discovered characteristic when it is partly enforced by the training loss. These issues leave the core claim supported only conditionally, pending additional validation experiments.","major_comments":[{"comment":"The pseudo-labeling step that drives the whole scene disentanglement is not validated against any ground-truth segmentation. K-Means clustering of superpixel appearance features and event correlation volumes is used to generate labels for training ClusterFormer via Eq. (1), but no IoU, accuracy, or other quantitative metric is reported against actual background/object regions. The t-SNE visualization in Fig. 7 is circular evidence because K-Means always produces separated clusters. Since the downstream 3D/4D Gaussian assignment, EKF tracking, and fusion in Eqs. (3)-(4) all depend on these masks, the manuscript must either report pseudo-label accuracy on datasets where ground-truth masks are available (e.g., synthetic Event-HyperNeRF), or ablate with ground-truth masks to show that the gains in Table 3 are not an artifact of incorrect segmentation. The number of clusters K is also never specified, and it is unclear how the fixed K=2 assumption handles scenes with multiple dynamic objects.","section":"Section 3.2, Eq. (1), Fig. 7"},{"comment":"The sentence 'we manually divide the regions of background and objects as candidate regions to match' is ambiguous and load-bearing for the claim that the disentanglement is automatic. If manually drawn region masks are used to extract or match the spatial and temporal features that feed the K-Means clustering, then the method is not fully automatic, and the comparison with fully automatic baselines in Tables 1 and 2 is not fair. If the manual division is used only for illustrative visualization, that should be stated explicitly and the entire pipeline should be shown to work without any manual input. As written, the reader cannot determine whether the reported gains come from the proposed appearance-motion clustering or from human annotations.","section":"Section 3.2, sentence on manual candidate regions"},{"comment":"The claim that 'we discover that Gaussian representations and event data share the consistent spatiotemporal characteristic' is not supported independently of the proposed loss. Eq. (2) explicitly enforces consistency between Gaussian color/deformation and event brightness/flow by minimizing the l1 difference, so any observed consistency after training is a consequence of this constraint, not a discovered prior. Fig. 4 illustrates the consistency but is presumably produced after or with the same loss. To support the conceptual claim in the abstract, the authors should provide evidence computed before optimization or on a held-out set without the consistency loss, or alternatively reframe the statement as a designed inductive bias rather than an empirical discovery.","section":"Section 3.3, Eq. (2), Fig. 4"},{"comment":"The comparison with frame-based baselines is not controlled for the effect of event-based deblurring. The proposed method first uses event data to restore potential blurry textures (Section 3.2, 'Referring to GEM'), whereas the 'w/ Deblur' baselines use a generic DeblurGAN without event information. The large margins in Tables 1 and 2 may therefore partly reflect the advantage of event-guided deblurring rather than the proposed spatiotemporal disentanglement. The authors should include a comparison with an event-based deblurring baseline, or ablate the event-deblur step within their own framework, to isolate the contribution of the disentanglement and fusion modules.","section":"Tables 1-2, Section 3.2"}],"minor_comments":[{"comment":"All ablation studies report single-run metrics without error bars or significance tests; reporting mean and standard deviation over multiple seeds would strengthen the claims, especially given the large reported margins.","section":"Table 3-6"},{"comment":"Several implementation details are missing: the number of K-Means clusters K, the architecture and hidden size of the shadow-overlap MLP in Eq. (3), the EKF noise parameters, and the exact form of the superpixel and correlation-volume features. These omissions hinder reproducibility.","section":"Section 3.4 and Eq. (1)"},{"comment":"The use of blur synthesis and frame extraction to simulate high-dynamic scenes from lower-dynamic data should be discussed more carefully, since the realism of this synthetic degradation directly affects the external validity of the evaluations.","section":"Section 4.1, Event-HyperNeRF and HD-CED"},{"comment":"Reference [32] contains a typo ('Clusterfomer' should be 'ClusterFormer'), and the repeated spelling 'TiNeuV ox' in Table 1 and the caption should be corrected.","section":"References"},{"comment":"The stated limitation about radial motion along the z-axis is honest, but it should be reflected in the abstract and conclusion, which currently claim general 'high-dynamic scene reconstruction' without noting this failure mode.","section":"Section 4.4, Limitation"}],"recommendation":"major_revision","confidential_remarks":"The core idea is potentially interesting, but the current evidence does not yet establish that the disentanglement mechanism is both reliable and the source of the reported gains. The authors should be encouraged to release code and the CoFED dataset, as the evaluation would then be reproducible and the pseudo-label validation could be checked independently. The limitation paragraph is a positive sign and should remain, but the scope claims should be adjusted accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nHere's my take on STD-GS. The paper is worth a serious referee. The core idea—disentangling background and object Gaussians using frame appearance and event motion features, then using event brightness and flow as pseudo-labels for object Gaussian color and deformation—is a genuinely new pipeline. The CoFED dataset, with pixel-aligned frame-event data from a coaxial device, is a useful addition to the field. The ablations in Tables 3–6 are internally consistent and support the value of each proposed component, including temporal tracking over frame interpolation.\n\nThe soft spots are real but addressable. First, the paper claims to 'discover' the consistency between Gaussian and event characteristics, but that consistency is imposed by the loss Lconsis in Eq. 2. That's circular phrasing; it should be reframed as a modeling choice, not a finding. Second, the DSD module's clustering pseudo-labels are never quantitatively validated against ground-truth segmentation. The t-SNE in Fig. 7 shows that clusters separate features, but K-Means always separates. The sentence in Sec. 3.2 about 'manually divide[ing] the regions of background and objects as candidate regions' is ambiguous—if human masks are used to generate or refine pseudo-labels, the method is not fully automatic and the gains may partially come from annotation. This needs a clear statement. Also, the cluster count K is not specified, which matters for scenes with multiple objects. Third, there are no error bars or significance tests, and part of the evaluation is on blur-synthesized versions of existing datasets, which is a reasonable practice but should be flagged. No code or data are released yet, which limits reproducibility.\n\nNone of these are load-bearing flaws; they're reporting and framing issues. The central claim—that spatiotemporal disentanglement improves high-dynamic scene reconstruction—holds up in the ablations, and the visual comparisons look convincing. The limitation about radial motion is honestly stated.\n\nI'd send this to peer review. The right reviewer is someone working on event-based 3D reconstruction; they'll want code and the clarified clustering supervision. If I were working in this area, I'd cite it for the pipeline and the dataset.\n\nBest,","headline":"Plausible event-guided disentanglement for dynamic scenes, with solid ablations and a useful new dataset, but the paper overstates its 'discovery' of Gaussian-event consistency and should clarify the clustering supervision and report error bars.","tokens_in":13417,"tokens_out":2744,"would_cite":true,"duration_ms":28128,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that separately modeling static background and dynamic objects as Gaussians, guided by frame and event data, reconstructs time-continuous high-dynamic scenes with sharper novel views.","keywords":["Gaussian splatting","event camera","dynamic scene reconstruction","spatiotemporal disentanglement","novel view synthesis","appearance-motion clustering","4D Gaussian","frame-event fusion"],"falsifier":"Capture or render a high-dynamic scene in which the moving object shares the background's color distribution and generates event flow similar to background edges, then check whether appearance-motion clustering still separates its superpixels into the object cluster; if the object Gaussians are not built, the novel view will show the object missing or blurred, and the claimed disentanglement would be falsified.","tokens_in":1710,"feed_emoji":"🎥","tokens_out":1743,"duration_ms":62772,"temperature":0.7,"pith_summary":"The paper tries to show that high-dynamic scene reconstruction fails when one Gaussian model is asked to represent both rigid background and deformed moving objects, and that separating the two before rendering fixes it. It claims that frame images supply dense spatial appearance while event cameras supply continuous temporal motion, so the two modalities can be combined to distinguish background from objects by clustering. The method then represents the background as 3D Gaussians and each dynamic object as 4D Gaussians centered on event-tracked points, and uses event data as pseudo-labels to constrain the object's color and deformation. A sympathetic reader would care because the reported numbers, for example 27.12 dB PSNR versus 20.60 dB for the best frame-based competitor on the HD-CED data, suggest that the disentanglement removes the blur and position shift that plague unified dynamic-scene methods.","feed_headline":"Event-guided splatting disentangles moving objects from scenes","feed_subtitle":"Frame plus event data rebuild time-continuous high-dynamic scenes with sharper novel views.","key_machinery":"The load-bearing object is a two-stage disentanglement. First, dynamic scene disentanglement constructs spatial appearance features via SLIC superpixels and temporal motion features via inner-product correlation volumes of the event stream, then clusters these features with K-Means to produce pseudo-labels; a ClusterFormer trained with a clustering loss $L_{clu}$ recursively updates cluster centers so that background and object features separate in feature space. Second, Gaussian representation fusion uses an extended Kalman filter on event correlation volumes to produce continuous tracking points, places 4D Gaussians on those points, and constrains them with a consistency loss $L_{consis}$ that matches Gaussian color to event brightness and Gaussian deformation to event optical flow. Background 3D Gaussians and object 4D Gaussians are combined by shadow-based weights and rendered with alpha-blending. The mechanism works because it converts the frame-event modality gap into explicit pseudo-labels, then uses event spatiotemporal structure as a training signal for the object's Gaussian parameters.","core_discovery":"The central claim is that a spatiotemporal-disentangled Gaussian splatting pipeline, called STD-GS, can reconstruct a time-continuous high-dynamic scene by modeling static background and dynamic objects in separate latent Gaussian spaces instead of one unified representation. Scene-level disentanglement uses a learnable appearance-motion clustering network: superpixel patches from frames are the spatial appearance cue, event correlation volumes are the temporal motion cue, K-Means supplies pseudo-labels, and a ClusterFormer refines the separation. Object-level disentanglement represents the object as 4D Gaussians centered on tracking points updated by an extended Kalman filter over event data, and ties Gaussian color and deformation to event brightness and flow through consistency losses. The two families of Gaussians are fused with shadow-aware weights and rendered by alpha-blending. The paper reports that this decomposition-fusion design outperforms unified and event-enhanced baselines on synthetic and real high-dynamic datasets, such as 27.12 dB PSNR versus 20.60 dB on HD-CED.","pith_inferences":["Editorial inference: the same appearance-motion clustering could be reused as a self-supervised segmentation step for other event-frame dynamic reconstruction and video editing pipelines, since it produces an explicit background-object mask rather than a latent-only split.","Editorial inference: the reported failure on radial z-axis motion suggests that a depth modality, or a stronger motion model, would be needed to extend the disentanglement to arbitrary 3D trajectories; the authors note this and propose LiDAR.","Editorial inference: if the clustering pseudo-labels are noisy, the consistency losses may reinforce the wrong object region; a testable extension would be to add uncertainty weighting or iterative label refinement during training.","Editorial inference: the method's benefit should transfer to deblurring and frame-interpolation benchmarks because the event-derived tracking directly addresses temporal discontinuity, though this is not demonstrated in the paper."],"forward_implications":["A single model need not carry both rigid and deformable motion: separating background and object Gaussians before fusion is enough to recover sharper novel views in high-dynamic scenes.","Event data can serve as pseudo-labels for Gaussian spatiotemporal structure, not just as a deblurring or interpolation aid, if the color-brightness and deformation-flow correspondences hold.","Time-continuous rendering of dynamic scenes is achievable by tracking objects in event streams and placing object Gaussians on the tracked points, replacing frame interpolation.","The joint reconstruction, clustering, and consistency objectives are trainable end-to-end, with the inference model reduced to a ClusterFormer and a Gaussian representation.","On the datasets used, the frame-plus-event input outperforms either modality alone, and the gap is larger in high-dynamic than in low-dynamic motion."],"supporting_citations":[{"why":"3D Gaussian splatting is the base representation and rendering backbone, supplying alpha-blending and the reconstruction loss that the method extends.","marker":"[2]"},{"why":"4D Gaussian splatting is the strongest frame-only dynamic baseline whose unified paradigm motivates the disentanglement argument.","marker":"[6]"},{"why":"E2GS is the closest event-enhanced Gaussian baseline that the method compares against and improves upon.","marker":"[30]"},{"why":"SLIC superpixels provide the spatial appearance features fed into the appearance-motion clustering.","marker":"[9]"},{"why":"The extended Kalman filter updates tracking points that serve as centers for the object 4D Gaussians.","marker":"[10]"},{"why":"HyperNeRF is the source dataset for the synthetic Event-HyperNeRF benchmark used in comparisons.","marker":"[22]"},{"why":"GEM supplies the event-based deblurring strategy used to restore blurry textures before clustering.","marker":"[31]"},{"why":"ClusterFormer is the deep clustering network that refines K-Means pseudo-labels for background-object separation.","marker":"[32]"},{"why":"v2e converts frames to realistic event data for the synthetic high-dynamic dataset.","marker":"[36]"},{"why":"CED is the real color event dataset used to build HD-CED for real-scene evaluation.","marker":"[37]"}],"fun_headline_variants":["Event-guided splatting untangles moving objects from scenes","Disentangled Gaussians plus events rebuild dynamic scenes","STD-GS: Frame-event Gaussians for high-dynamic reconstruction","Spatiotemporal disentanglement lifts dynamic scene rendering"],"cache_read_input_tokens":15488,"weakest_assumption_plain":"The pipeline assumes that K-Means clustering of superpixel appearance and event correlation volumes produces reliable background-object pseudo-labels that a ClusterFormer can learn; if the clustering is wrong, the 3D/4D Gaussian decomposition is built on incorrect regions and the disentanglement benefit collapses.","fun_headline_variants_meta":{"raw":{"variants":["Event-guided splatting untangles moving objects from scenes","Disentangled Gaussians plus events rebuild dynamic scenes","STD-GS: Frame-event Gaussians for high-dynamic reconstruction","Spatiotemporal disentanglement lifts dynamic scene rendering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00016,"raw_usage":{"total_tokens":1262,"prompt_tokens":1003,"completion_tokens":259,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":191}},"tokens_in":619,"tokens_out":259,"duration_ms":3050,"temperature":1.0,"reasoning_tokens":191,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:47:20.590764+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Capture or render a high-dynamic scene in which the moving object shares the background's color distribution and generates event flow similar to background edges, then check whether appearance-motion clustering still separates its superpixels into the object cluster; if the object Gaussians are not built, the novel view will show the object missing or blurred, and the claimed disentanglement would be falsified.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"3D Gaussian splatting is the base representation and rendering backbone, supplying alpha-blending and the reconstruction loss that the method extends."},{"cited_title":"4d gaussian splatting for real-time dynamic scene rendering","cited_arxiv_id":null,"evidence_quote":"4D Gaussian splatting is the strongest frame-only dynamic baseline whose unified paradigm motivates the disentanglement argument."},{"cited_title":"E2gs: Event enhanced gaussian splatting","cited_arxiv_id":null,"evidence_quote":"E2GS is the closest event-enhanced Gaussian baseline that the method compares against and improves upon."},{"cited_title":"Slic superpixels compared to state-of-the-art superpixel methods","cited_arxiv_id":null,"evidence_quote":"SLIC superpixels provide the spatial appearance features fed into the appearance-motion clustering."},{"cited_title":"Robust extended kalman filtering","cited_arxiv_id":null,"evidence_quote":"The extended Kalman filter updates tracking points that serve as centers for the object 4D Gaussians."},{"cited_title":"Generalizing event-based motion deblurring in real-world scenarios","cited_arxiv_id":null,"evidence_quote":"GEM supplies the event-based deblurring strategy used to restore blurry textures before clustering."},{"cited_title":"Clusterfomer: clustering as a universal visual learner","cited_arxiv_id":null,"evidence_quote":"ClusterFormer is the deep clustering network that refines K-Means pseudo-labels for background-object separation."},{"cited_title":"v2e: From video frames to realistic dvs events","cited_arxiv_id":null,"evidence_quote":"v2e converts frames to realistic event data for the synthetic high-dynamic dataset."},{"cited_title":"Ced: Color event camera dataset","cited_arxiv_id":null,"evidence_quote":"CED is the real color event dataset used to build HD-CED for real-scene evaluation."}],"review_version":1}