{"id":"32bb25c9-8955-44b9-9e6b-ee37428ed12f","arxiv_id":"2411.15582","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A learnable motion embedding and two-level deformation module improve dynamic street-scene rendering for several Gaussian splatting baselines.","lead":"This paper describes a plug-and-play addition to street-scene reconstruction methods that helps them capture different motion speeds of vehicles and pedestrians. It reports quality improvements on public driving datasets, a step toward more realistic simulators for autonomous driving.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be confounded by changed training recipe (LiDAR initialization + sky loss), so EMD's contribution is untested.","rationale":"The reader's weakest assumption was that the coarse/fine motion decomposition is not enforced by any loss, so the mechanism may not be doing the claimed work. That is a real conceptual concern, but the more immediately load-bearing issue is attribution: the paper's experiments do not include a control where the shared training changes (LiDAR-prior initialization and sky supervision) are applied to the baseline without EMD. The ablations in Table 6 remove one EMD component at a time from the full model, but they never show what the baseline achieves once it is given the same initialization and loss terms. If the shared changes account for most of the PSNR gain, then even a well-enforced motion decomposition is not necessary for the reported improvement, and the central claim that EMD is the cause of the gains collapses. This is an internal-consistency issue: the paper's own experimental design cannot distinguish EMD's contribution from the surrounding recipe changes. The proposed test directly fills that gap. It does not require rejecting the paper; it only means the conditional verdict stays conditional until such a control is reported. I therefore keep the reader's CONDITIONAL verdict and classify my agreement as partial, since the reader identified a related but different weakest point.","tokens_in":16098,"tokens_out":7225,"duration_ms":67756,"concrete_test":"Retrain S3Gaussian under the paper's modified training recipe (LiDAR initialization, Lsky and shared losses, same 50k iterations) but with the EMD-specific components removed: no per-Gaussian embedding, no dual-scale coarse/fine head, and a single deformation head of matched capacity. If the resulting full-image PSNR on Waymo-D32 is close to the reported 32.50 rather than the original 30.69, EMD is not the source of the reported gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To establish EMD as a drop-in performance booster, the comparison must hold everything except EMD fixed. The paper changes more than EMD when moving from S3Gaussian to S3Gaussian+Ours. Supplementary C.1 adds LiDAR-prior initialization (replacing SFM points) and introduces Lsky, a sky supervision loss, into the S3Gaussian loss. Table 6 ablates EMD components only from the full model, so the full model's gains over the published S3Gaussian baseline could be caused by these shared changes rather than by per-Gaussian embeddings or dual-scale deformation. If that is the case, the central claim that 'previous methods with EMD exhibit better reconstruction quality' is not established, even if the mechanism story were sound. The reported +1.81 PSNR over S3Gaussian in Table 1 and the +0.54 over DeSiRe-GS in Table 2 are therefore not cleanly attributable to EMD without a matched baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Explicit Motion Decomposition (EMD), a module for street Gaussian Splatting that augments Gaussian primitives with per-Gaussian learnable motion embeddings, uses an adaptive temporal embedding matrix, and introduces a dual-scale deformation framework with coarse and fine decoders. EMD is integrated into self-supervised baselines S3Gaussian and DeSiRe-GS, and extended to supervised methods StreetGaussian and OmniRe. Experiments on Waymo and KITTI report improved scene reconstruction and novel view synthesis in several settings, along with novel trajectory synthesis evaluated by FID. The paper claims that EMD is a plug-and-play module that achieves state-of-the-art self-supervised novel view synthesis performance.","tokens_in":16306,"tokens_out":3526,"duration_ms":34372,"significance":"If the attribution were clean, the contribution would be practically valuable: a drop-in module that improves multiple existing street-Gaussian pipelines, with code release, comparisons across four base methods and two datasets, ablation studies, and parameter sensitivity analysis. The manuscript also introduces a useful evaluation component in novel trajectory synthesis. However, the central attribution is not established as reported, because the S3Gaussian+Ours comparison changes more than EMD, and the headline claim of state-of-the-art novel view synthesis is contradicted by the paper's own Table 1. These issues are fixable with matched baselines and corrected claims, so the work has potential but needs substantial revision.","major_comments":[{"comment":"The central comparison between S3Gaussian and S3Gaussian+Ours does not isolate EMD. Supplementary C.1 states that the proposed method replaces the original SFM initialization with LiDAR-prior initialization and adds a sky supervision loss Lsky into the original S3Gaussian loss. Because these changes are applied only to the EMD-augmented model and not to the reported baseline, the +1.81 PSNR improvement in Table 1 (and the gains in Table 2) cannot be attributed to EMD. The authors should re-run the original S3Gaussian and DeSiRe-GS with the same LiDAR initialization, sky loss, and any other shared training-recipe changes, and then add EMD on top; this matched comparison is essential to support the claim that 'previous methods with EMD exhibit better reconstruction quality.'","section":"§4.2.1, Supplementary C.1, Tables 1–2"},{"comment":"The full-image novel view synthesis PSNR for S3Gaussian+Ours is 26.55, which is lower than S3Gaussian's 26.62, while the abstract and contributions claim state-of-the-art novel view synthesis performance in self-supervised settings. This directly undercuts the blanket NVS claim. The authors should either report an appropriate aggregate or scenario-specific metric that supports the claim, restrict the claim to scene reconstruction and selected object regions, or provide additional experiments on novel view synthesis where the improvement is consistent.","section":"§5.2.1, Table 1"},{"comment":"The mechanism story that coarse deformation captures fast global motion while fine deformation captures slow local deformation, and that per-Gaussian embeddings encode motion speed, is not enforced or tested. Eq. (11) simply sums the two deformation outputs, and Eq. (15) only encourages neighboring Gaussians to have similar embeddings; no loss separates motion scales or ties embeddings to speed. The ablations in Table 6 show that removing the coarse branch causes a large drop and removing the fine branch hurts LPIPS, but this does not demonstrate specialization. The authors should add a direct test, such as measuring displacement magnitudes of coarse vs fine branches per object class or visualizing embedding clustering by speed, or else soften the claims to describe a general increase in deformation network capacity.","section":"§4.1.2, Eq. (11), §4.2.1, Eq. (15), Table 6"},{"comment":"Several reported improvements are small (e.g., DeSiRe-GS+Ours vs DeSiRe-GS is +0.54 PSNR on Waymo scene reconstruction, +0.16 on novel view synthesis, and +0.18/+0.17 on KITTI), and no error bars, per-scene standard deviations, or significance tests are provided. Given that the comparisons may include training-recipe confounds, the numerical claims are not yet robust; reporting per-scene variance across the small scene subsets would materially strengthen the paper.","section":"Tables 1–2, 4–6"}],"minor_comments":[{"comment":"The table header contains a typo: 'StreetGaussin' should be 'StreetGaussian'.","section":"Table 3"},{"comment":"The caption contains placeholder question marks ('?????') in place of variant descriptions; these should be replaced with explicit labels.","section":"Figure 6"},{"comment":"The text says 'We also report FPS to access inference speed'; 'access' should be 'assess'.","section":"§5.1"},{"comment":"The notation KNN_i;d is not defined precisely; the text says 'KNN means the k-nearest-neighbors algorithm' but does not clarify how the hyperparameter d in Eq. (15) relates to the number of neighbors k.","section":"Eq. (15)"},{"comment":"The reference labeled DeformGS [63] is cited in Table 4 but the bibliography entry describes 'Deformable 3D Gaussians'; please unify the name and citation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope, and the proposed module is plausible and potentially useful. The main risk is attribution: the EMD-vs-baseline comparisons change LiDAR initialization and the sky loss alongside EMD, so the headline gains may not come from the proposed components. The novel-view-synthesis claim also conflicts with Table 1. I do not see the flaws as unfixable; matched-baseline experiments and a corrected claim would be sufficient for reconsideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper reports solid but not cleanly attributable gains. The EMD module—per-Gaussian latent embeddings, progressive temporal sampling, and a coarse/fine deformation split—is a reasonable engineering extension of existing dynamic Gaussian work, and the experiments cover four baselines and two datasets. But because the S3Gaussian comparison changes more than EMD (LiDAR-prior init and a sky loss are added in the supplementary), the +1.81 PSNR in Table 1 cannot be credited solely to the module.\n\nWhat is actually new: the targeted integration and the supervision-agnostic design. Previous self-supervised street GS methods don't model per-object motion speed; EMD adds a lightweight, compatible module that improves reconstruction and vehicle metrics on Waymo and KITTI. The ablations (Table 6) show each component contributes, with coarse deformation being the biggest lever. That is useful evidence for the street rendering community.\n\nSoft spots, in order of severity. First, the confound. Supplementary C.1 changes initialization and adds Lsky to the baseline loss; the published S3Gaussian numbers in Table 1 don't have those changes, so the comparison is not controlled. The authors should run S3Gaussian with the same init and sky loss, with and without EMD, or at least ablate those factors. Second, the abstract's 'state-of-the-art novel view synthesis' claim is contradicted by Table 1: full-image novel-view PSNR is 26.55 vs 26.62 for S3Gaussian. The gains are in SSIM/LPIPS and vehicle PSNR, which are meaningful, but the claim should be narrowed. Third, no error bars or significance tests, and code isn't actually released despite the abstract's 'code is available' pointer. Fourth, the mechanistic story—that coarse handles fast/global and fine handles slow/local—is asserted, not enforced; Eq. (11) just sums the two outputs, so the gains could come from extra capacity or better optimization. That's a minor criticism because the ablation does show both scales matter.\n\nBottom line: the paper deserves a serious referee. The core integration is potentially useful for driving simulators, but the central comparative claim needs a matched-baseline experiment before the numbers can be trusted. I'd send it to review with a request for controlled comparisons and a toned-down abstract.","headline":"A useful plug-and-play motion module for street Gaussian splatting, but the headline gains are partly confounded by an unablated change in the training recipe.","tokens_in":16858,"tokens_out":1889,"would_cite":false,"duration_ms":16608,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding motion embeddings sharpens street-scene Gaussian rendering","keywords":["Gaussian splatting","street scene reconstruction","self-supervised learning","motion decomposition","novel view synthesis","autonomous driving","deformation field"],"falsifier":"A direct test: give both deformation branches identical capacity and input features, or swap their order; if rendering quality stays essentially unchanged, the claimed coarse-fast versus fine-slow decomposition is not what drives the gains. A second check: read out the learned embedding $z_k$ for Gaussians on pedestrians versus vehicles and test whether the embedding space separates by measured object speed.","tokens_in":15876,"feed_emoji":"🚗","tokens_out":3840,"duration_ms":32847,"temperature":0.7,"pith_summary":"The paper argues that street-scene Gaussian splatting reconstructions look blurry because existing methods treat dynamic objects as one binary class and ignore the continuous range of motion speeds: a pedestrian and a vehicle are modeled with the same deformation machinery. To fix this, it introduces Explicit Motion Decomposition (EMD), a plug-and-play module that attaches a learnable motion embedding to each Gaussian primitive and splits deformation into a coarse branch for large-scale motions and a fine branch for local details. Adding EMD to two self-supervised street-splatting baselines raises reconstruction quality on Waymo and KITTI benchmarks, and the same module extends to box-supervised pipelines by refining tracked poses and SMPL body parameters. If the claim holds, motion modeling can be improved without changing the underlying scene representation or requiring new annotations.","feed_headline":"Adding motion embeddings sharpens street-scene Gaussian rendering","feed_subtitle":"A plug-in module splits deformation into coarse and fine stages, lifting reconstruction quality on Waymo and KITTI.","key_machinery":"The load-bearing object is the dual-scale deformation network combined with per-Gaussian learnable embeddings. Each Gaussian carries a latent vector $z_k \\in \\mathbb{R}^{32}$; a shared temporal embedding matrix $W$ is sampled through progressive bilinear interpolation so that early training sees coarse time samples and later training sees finer ones; the aggregated feature $\\mathcal{F}_{aggr}(\\mu,t) = [\\mathcal{F}_{pos}(\\mu), \\mathcal{F}_{temp}(t), z_k]$ feeds a coarse multi-head decoder and a fine multi-head decoder in series. The coarse decoder handles large displacements such as vehicle translation, the fine decoder receives $\\mu + \\Delta\\mu_{coarse}$ and handles local detail such as articulated motion. Ablations show removing the coarse branch drops full-image PSNR by 3.10 dB, while removing the fine branch worsens perceptual quality.","core_discovery":"On its own terms, the paper's central claim is that explicit motion modeling is the missing ingredient in current street Gaussian splatting. The proposed EMD module gives each Gaussian a learnable embedding $z_k$ that encodes its individual motion characteristics, combines spatial, temporal, and per-Gaussian features into a motion-aware encoding, and predicts deformations in two stages: $\\mathcal{D}_{coarse}$ for global, fast motions such as vehicle translation and $\\mathcal{D}_{fine}$ for local, articulated deformation, with the final deformed position computed as $\\mu_t = \\mu + \\Delta\\mu_{coarse} + \\Delta\\mu_{fine}$. Integrated into S3Gaussian and DeSiRe-GS, EMD improves PSNR, SSIM, and LPIPS on Waymo and KITTI; integrated into StreetGaussian and OmniRe, it improves supervised novel-view synthesis. The paper further reports that EMD substantially lowers FID on novel trajectory synthesis, arguing that better motion modeling directly benefits the lane-change simulation use case.","pith_inferences":["If the learned embeddings genuinely encode motion speed, they might transfer across scenes: an embedding space pretrained to separate vehicle-like from pedestrian-like motion could accelerate training on new street datasets.","The paper does not enforce that the coarse branch handles fast motion and the fine branch slow motion; a plausible alternative explanation for the gains is simply that two-stage deformation adds network capacity. Swapping or equalizing the two branches would test this.","The FID improvement on shifted trajectories implies that motion modeling quality can be measured without ground-truth novel views, which could make FID on shifted trajectories a standard evaluation for driving simulators.","The limitation section notes lighting is unmodeled; the same embedding-plus-dual-scale idea could naturally extend to lighting or weather-dependent appearance."],"forward_implications":["Any self-supervised street Gaussian splatting method can adopt EMD as a drop-in module and expect better reconstruction, especially on fast-moving vehicles.","Novel trajectory synthesis for lane-change simulation improves, with FID dropping from 83.48 to 45.11 at 0.5 m offset on Waymo.","Supervised pipelines with tracked boxes and SMPL bodies also benefit: StreetGaussian + EMD and OmniRe + EMD improve novel-view PSNR and SSIM.","The progressive temporal embedding schedule suggests a general coarse-to-fine training strategy for deformation networks, not limited to street scenes."],"supporting_citations":[{"why":"Supplies the S3Gaussian self-supervised baseline whose HexPlane representation and multi-head decoder EMD augments.","marker":"[16]"},{"why":"Supplies the DeSiRe-GS self-supervised baseline that EMD enhances with learnable embeddings and dual-scale deformation.","marker":"[38]"},{"why":"Supplies the StreetGaussian supervised baseline with tracked vehicle poses and object Gaussians that EMD refines.","marker":"[60]"},{"why":"Supplies the OmniRe supervised baseline with SMPL body modeling that EMD extends to non-rigid human motion.","marker":"[6]"},{"why":"Provides the base 3D Gaussian Splatting representation and rendering pipeline all compared methods build on.","marker":"[19]"},{"why":"Supplies the PVG periodic-vibration Gaussian framework that DeSiRe-GS follows and EMD modifies.","marker":"[5]"},{"why":"Defines the Waymo D32 benchmark split and serves as a baseline for self-supervised street reconstruction.","marker":"[61]"}],"fun_headline_variants":["Explicit motion embeddings improve street Gaussian splatting","Per-Gaussian motion encoding boosts street rendering quality","Plug-in motion module sharpens dynamic street Gaussian scenes","Split motion into coarse and fine to sharpen street Gaussians","Learnable motion embeddings lift street Gaussian reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the coarse/fine split actually separates fast global motion from slow local deformation and that the learned per-Gaussian embedding encodes motion speed; nothing in the training loss forces either property, since the loss only sums the two deformation outputs and encourages neighboring Gaussians to have similar embeddings.","fun_headline_variants_meta":{"raw":{"variants":["Explicit motion embeddings improve street Gaussian splatting","Per-Gaussian motion encoding boosts street rendering quality","Plug-in motion module sharpens dynamic street Gaussian scenes","Split motion into coarse and fine to sharpen street Gaussians","Learnable motion embeddings lift street Gaussian reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1337,"prompt_tokens":992,"completion_tokens":345,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":271}},"tokens_in":608,"tokens_out":345,"duration_ms":3347,"temperature":1.0,"reasoning_tokens":271,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:08:53.208769+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test: give both deformation branches identical capacity and input features, or swap their order; if rendering quality stays essentially unchanged, the claimed coarse-fast versus fine-slow decomposition is not what drives the gains. A second check: read out the learned embedding $z_k$ for Gaussians on pedestrians versus vehicles and test whether the embedding space separates by measured object speed.","supporting_citations":[{"cited_title":"Street gaussians: Modeling dynamic urban scenes with gaussian splatting","cited_arxiv_id":null,"evidence_quote":"Supplies the StreetGaussian supervised baseline with tracked vehicle poses and object Gaussians that EMD refines."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the base 3D Gaussian Splatting representation and rendering pipeline all compared methods build on."}],"review_version":1}