{"id":"fac66390-8266-490d-a27d-0750f9e791f3","arxiv_id":"2411.15482","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SplatFlow learns 4D Gaussian scene representations inside a neural motion flow field, achieving state-of-the-art novel view synthesis on Waymo and KITTI without tracked 3D boxes.","lead":"SplatFlow is a new method for reconstructing dynamic driving scenes without 3D bounding box labels. It combines 4D Gaussian splatting with a learned neural motion flow field to separate moving objects from the static background on its own.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Static/dynamic decomposition depends on an unspecified motion-flow threshold; without sensitivity analysis or flow accuracy checks, the box-free SOTA claim is not yet fully supported.","rationale":"The reader's weakest assumption identifies exactly the load-bearing point: the static/dynamic split is produced by an unspecified threshold on learned 3D motion flow, and the flow itself derives from pseudo scene flow labels generated by an external method. This is the right concern because the dynamic mask is the gate that decides which Gaussians are 4D and which are static; any error there directly undermines the 'without tracked 3D bounding boxes' claim. I examined whether another issue—such as the reliance on SEA-RAFT optical flow distillation or the 'self-supervised' label—was more central. The optical flow distillation is an auxiliary loss and is ablated positively, so it is not the weakest point. The self-supervision terminology is weakened by pseudo labels, but the core technical premise remains the thresholded decomposition. The paper gives no value for the threshold, no sensitivity analysis, and no flow accuracy metrics, so the reported SOTA numbers cannot be independently reproduced or assessed for robustness. The ablation in Table 5 shows the NMFF prior contributes about 1.3 dB, but it does not test threshold sensitivity. Therefore the verdict stays CONDITIONAL: the method is plausible and the ablations are encouraging, but the central decomposition needs a specified threshold, sensitivity analysis, or code release before the claim can be accepted as definitive.","tokens_in":14193,"tokens_out":4984,"duration_ms":48714,"concrete_test":"Sweep the dynamic-mask threshold over a meaningful range (e.g., 0.01–1.0 m of predicted 3D flow magnitude) for at least one Waymo scene (e.g., Seg. 2259) and recompute the dynamic-only metrics (Table 2 style). If PSNR* or SSIM* changes by more than about 5% relative across the range, or if the optimal threshold differs by more than 2x between scenes, the decomposition is too threshold-sensitive to support a general box-free SOTA claim. Report the chosen threshold value alongside the result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—SOTA dynamic urban reconstruction without tracked 3D boxes—rests on the NMFF-based decomposition of LiDAR points into static and dynamic sets. Section 3.4 states that 'a 3D dynamic mask to separate static and dynamic points' is produced 'by applying a threshold on each point's 3D motion flow', but no threshold value, selection criterion, or sensitivity analysis is given anywhere in the paper or supplement. The NMFF is pretrained (Supplement Section 6) on pseudo scene flow labels from [14] using Chamfer distance on raw LiDAR; such correspondences are known to be unreliable for fast-moving, distant, or occluded objects and for repetitive geometry. If the threshold is set too high, moving vehicles are merged into the static 3D Gaussians and the 4D representation cannot recover them; if set too low, static background is fractured into spurious dynamic Gaussians. Since the dynamic mask determines the initialization and time-varying behavior of every Gaussian, the reported gains in Tables 1–4 are contingent on this unspecified hyperparameter. Additionally, no error bars or per-scene threshold values are reported, so it is unclear whether the SOTA numbers reflect a robust choice or a favorable one.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"SplatFlow proposes a self-supervised dynamic Gaussian splatting framework for autonomous driving scenes that avoids tracked 3D bounding boxes. The method introduces a Neural Motion Flow Field (NMFF), pretrained on temporally consecutive LiDAR point clouds with Chamfer distance, to predict 3D motion flow. A threshold on this flow decomposes points into static and dynamic sets, which seed static 3D Gaussians and dynamic 4D Gaussians, respectively. During Gaussian splatting, the NMFF is jointly optimized to propagate 4D Gaussians across time, and optical flow from SEA-RAFT is distilled into the representation. The paper reports state-of-the-art results on the Waymo Open Dataset and KITTI for image reconstruction and novel view synthesis, including ablations and runtime comparisons.","tokens_in":14409,"tokens_out":3581,"duration_ms":32747,"significance":"If the reported results hold, SplatFlow is a meaningful step toward scalable dynamic scene reconstruction for autonomous driving simulation, as it removes the dependence on expensive 3D bounding-box annotations that most leading dynamic urban Gaussian splatting methods require. The unified NMFF representation that simultaneously drives static/dynamic decomposition and cross-time 4D Gaussian correspondence is a well-conceived design. The paper provides clear ablations showing the contribution of each component, reports real-time inference FPS, and backs the claims with extensive qualitative visualizations. These strengths are, however, tempered by the absence of a code release and by several specification gaps that currently prevent full reproducibility and may affect the robustness of the SOTA claim.","major_comments":[{"comment":"The static/dynamic decomposition is obtained 'by applying a threshold on each point's 3D motion flow' (Section 3.4), but the threshold value, its selection criterion, and any sensitivity analysis are never reported in the manuscript or supplement. This threshold determines which LiDAR points seed 4D dynamic Gaussians and which seed static 3D Gaussians, so the entire 4D representation—and hence the NVS results in Tables 1–4—is contingent on an unspecified hyperparameter. Please provide the threshold value and an ablation over a range of thresholds (e.g., NVS metrics as a function of threshold), or a principled criterion (e.g., a percentile of flow magnitude) that makes the choice canonical.","section":"Section 3.4"},{"comment":"The reported average metrics have no error bars or per-scene standard deviations. For the central SOTA claim, the differences against the strongest baseline PVG are modest in some cells (e.g., Table 3 KITTI image reconstruction SSIM 0.943 vs 0.937; LPIPS 0.057 vs 0.070), and without variance estimates it is impossible to judge whether these differences are statistically meaningful or driven by a few favorable segments. Please provide per-scene result tables and/or standard deviations across scenes and runs.","section":"Tables 1, 3, 4"},{"comment":"The NMFF is pretrained on pseudo scene flow labels from [14] generated with Chamfer distance on raw LiDAR, but the paper never evaluates the accuracy of the learned 3D motion flow against any reference or manual annotations, nor does it analyze failure modes for fast-moving, distant, or occluded objects. Because the dynamic mask directly inherits errors from this prior, the box-free claim and the overall robustness depend on NMFF flow quality. Please provide a quantitative evaluation of NMFF's flow (e.g., end-point error on a validation subset with known object motions) or a sensitivity test where pseudo labels are perturbed (e.g., by noise) and the resulting NVS metrics are reported.","section":"Supplement Section 6"}],"minor_comments":[{"comment":"The word 'Levaraging' should be 'Leveraging'.","section":"Introduction, page 2"},{"comment":"The captions contain typos: 'SlatFlow' and 'SlpatFlow' should be 'SplatFlow'.","section":"Figure 11, Figure 15, Figure 17"},{"comment":"The segment identifiers are partially redacted with ellipses (e.g., 'Seg. 1058...'); please provide the full segment IDs for reproducibility.","section":"Table 2"},{"comment":"The terms 'mid-timestamp reference frame' and 'step-by-step propagation' are not precisely defined; please specify how the reference timestamp is selected and how aggregation of 4D Gaussians across timestamps is performed.","section":"Section 3.4"},{"comment":"Equation (10) uses µ′ for projected centers, but the rendering equations in Eq. (3) use depth z_i without linking the two notations; clarify the projection and depth computation for the flow rendering.","section":"Section 3.5"},{"comment":"The runtime claim of 40–44 FPS is based on 'quantization and pruning optimization', but no details of the quantization method or pruning ratio are given; please specify these to support the real-time claim.","section":"Supplement Section 12"}],"recommendation":"major_revision","confidential_remarks":"The load-bearing issue is indeed the unspecified static/dynamic threshold and the lack of robustness analysis around it. The paper's self-supervision framing is also somewhat broader than the title suggests, since it relies on pseudo-labels from an external scene flow method and on SEA-RAFT optical flow during training; this is worth clarifying but is not a blocker. If the authors supply the missing threshold details and a sensitivity study, this paper would likely become acceptable for a top-tier venue. I would encourage asking for the per-scene results as well."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a credible systems-and-method paper, not a paradigm shift. The claim that matters—box-free dynamic urban reconstruction with 4D Gaussians—is supported by internally consistent tables and a clean ablation, with one genuinely important missing number: the threshold that separates static from dynamic points is never stated.\n\nWhat is new: the Neural Motion Flow Field (NMFF) pretrained on LiDAR with Chamfer distance provides motion priors for Gaussian initialization; dynamic objects become 4D Gaussians, static background stays 3D; SEA-RAFT optical flow is distilled into the 4D representation. That combination is not in PVG, EmerNeRF, or StreetGaussian. The paper handles the natural comparison to PVG carefully—per-scene dynamic-only metrics in Table 2 are useful, and the ablations show each component contributes. That is real work and worth crediting.\n\nSoft spots, in proportion. The threshold in Section 3.4 really is load-bearing. It decides which LiDAR points become 4D Gaussians, so an unspecified value with no sensitivity analysis leaves the SOTA claim contingent. I agree with the stress-test note here. The NMFF prior is also pretrained on pseudo scene flow labels from [14] over raw LiDAR; fast, distant, or occluded objects are exactly where those labels are weak. The paper does not quantify how many moving objects the mask misses. Baseline numbers are reproduced from other papers, there are no error bars, and no code is released. None of this is evidence of a broken method; it is evidence that the current SOTA statement is one careful hyperparameter choice away from being verifiable.\n\nAlso, 'self-supervised' is doing some work: no tracked 3D boxes is true, but the model consumes pseudo optical flow from SEA-RAFT, a sky mask from SegFormer, and pretrained flow labels. That is a teacher-student setup, not circular reasoning, but it should be described as 'annotation-free' rather than fully self-supervised.\n\nBottom line: for readers working on dynamic driving reconstruction, Gaussian splatting, or closed-loop simulation, this paper is worth reading. I would send it to a serious referee, and I would ask for the threshold value, sensitivity analysis, code release, and error bars in revision.","headline":"Solid box-free dynamic Gaussian splatting with a real gap around the static/dynamic threshold; worth refereeing.","tokens_in":100,"tokens_out":1919,"would_cite":true,"duration_ms":35602,"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":"SplatFlow claims that a self-supervised 4D Gaussian representation inside a neural motion flow field reconstructs and renders dynamic urban scenes without tracked 3D bounding boxes.","keywords":["Dynamic Gaussian Splatting","Neural Motion Flow Field","Self-supervised scene decomposition","Novel view synthesis","Autonomous driving","LiDAR-camera fusion","Optical flow distillation","4D scene representation"],"falsifier":"Take a scene containing a small, mostly occluded vehicle moving in 3D but slowly in image space; render it with SplatFlow, then render the same scene after forcing the dynamic mask from ground-truth 3D bounding boxes. If the flow-based mask misses the vehicle, the dynamic-region PSNR will drop sharply relative to the box-forced version, while static metrics stay similar.","tokens_in":13981,"feed_emoji":"🚗","tokens_out":7465,"duration_ms":60729,"temperature":0.7,"pith_summary":"The paper sets out to show that dynamic urban driving scenes can be reconstructed and rendered from novel viewpoints without any tracked 3D bounding boxes, the supervision that most prior dynamic Gaussian methods require. To do this it introduces SplatFlow, which fuses time-varying 4D Gaussian primitives with a Neural Motion Flow Field (NMFF) that predicts where LiDAR points and Gaussians move between frames. A threshold on the predicted 3D motion separates static background (3D Gaussians) from moving objects (4D Gaussians), and the same field warps dynamic Gaussians to the target timestamp before rasterization. On Waymo and KITTI the paper reports higher PSNR, SSIM, and LPIPS than prior baselines, including box-supervised ones, for both image reconstruction and novel view synthesis. A sympathetic reader would take the claim to be that motion flow learned from LiDAR plus self-supervision can replace object-level annotation as the key prior for scalable driving-scene simulation.","feed_headline":"Self-supervised road-scene splatting beats box-labeled baselines","feed_subtitle":"Four-dimensional Gaussians plus a learned motion flow field split moving objects from background without tracked boxes.","key_machinery":"The load-bearing mechanism is the Neural Motion Flow Field (NMFF), a temporal sequence of MLP fields $\\phi_{t_1:t_2}$ that map any query point at time $t_1$ to its 3D displacement and motion angle at time $t_2$. Each field is eight ReLU-MLP stacks. NMFF does two jobs: it produces a 3D dynamic mask by thresholding per-point flow magnitude, deciding which LiDAR points initialize static 3D Gaussians versus dynamic 4D Gaussians, and it warps aggregated 4D Gaussians across timestamps so that at render time all Gaussians are brought to the target time and splatted. The authors find that rendering loss alone does not sufficiently constrain motion, so they pretrain the field on LiDAR geometry with bidirectional Chamfer distance and add optical-flow distillation from a 2D foundation model to inject motion knowledge from images.","core_discovery":"SplatFlow's central claim is that a 4D Gaussian splatting representation, when coupled to a neural motion flow field, can learn dynamic urban scenes without tracking or 3D bounding-box annotations. The motion field predicts where each LiDAR point and Gaussian moves between consecutive timestamps; the authors pretrain it on LiDAR point clouds with a bidirectional Chamfer distance, then jointly optimize it with Gaussian attributes on image and depth losses, and distill optical flow from a 2D foundation model. Static content is stored as 3D Gaussians and dynamic content as 4D Gaussians that are warped to a target time before rasterization. On Waymo and KITTI, the paper reports higher PSNR, SSIM, and LPIPS than prior baselines, including box-supervised ones, for both image reconstruction and novel view synthesis, with the strongest gains on dynamic regions.","pith_inferences":["Not pursued in the paper: the same NMFF motion prior could be reused as a warm start for other 4D representations, such as deformable radiance fields or motion forecasting modules, since it already outputs dense 3D point trajectories.","Not pursued in the paper: replacing the unspecified threshold with a confidence-weighted soft mask could make the static/dynamic decomposition robust without per-scene threshold tuning; this is a testable variant of the method as described.","Not pursued in the paper: because SplatFlow renders dense optical flow and depth, the same pipeline could be turned around to generate pseudo-labels for training scene-flow or optical-flow networks in driving data, where such labels are scarce."],"forward_implications":["Dynamic urban scenes can be reconstructed and rendered at roughly 40-44 FPS without any tracked 3D bounding boxes, using only calibrated cameras and LiDAR plus pretrained motion and optical-flow pseudo-labels.","The learned motion field enables novel-view synthesis at arbitrary timestamps, including ego trajectories shifted 0.5 meters from the recorded one, with rendered RGB, depth, and optical flow.","Because the dynamic mask comes from per-point motion magnitude rather than detectors, the method can scale to unlabeled in-the-wild driving data, which the authors identify as the main scalability bottleneck it removes.","Ablations show that each component, LiDAR flow pretraining, NMFF optimization with 4D Gaussians, and optical-flow distillation, contributes to the final rendering quality, especially inside dynamic regions."],"supporting_citations":[{"why":"Primary self-supervised Gaussian baseline and evaluation protocol; SplatFlow is measured against PVG and inherits its static and sky Gaussian handling.","marker":"[1]"},{"why":"Base 3D Gaussian splatting representation and differentiable rasterizer that the 4D formulation extends.","marker":"[7]"},{"why":"Supplies pseudo scene-flow labels from point clouds used to pretrain NMFF with bidirectional Chamfer distance.","marker":"[14]"},{"why":"Provides the 2D optical-flow pseudo-labels distilled into the 4D space-time representation during training.","marker":"[21]"},{"why":"Box-supervised Gaussian baseline; its reliance on tracked 3D bounding boxes is the scalability problem SplatFlow removes.","marker":"[25]"},{"why":"Self-supervised NeRF baseline that also decomposes scenes with flow; supplies the dynamic-mask evaluation protocol for dynamic-region metrics.","marker":"[26]"}],"fun_headline_variants":["Self-supervised 4D Gaussians best box-labeled splatting","No boxes needed: Neural motion flow drives dynamic splatting","Dynamic urban scenes without tracking via self-supervised 4D Gaussians","Motion flow field turns 4D Gaussians self-supervised on road scenes","SplatFlow: Label-free dynamic splatting outdoes supervised baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The static/dynamic split rests on a fixed threshold applied to the learned 3D motion flow; if the flow field underestimates the motion of small, distant, or occluded objects, those objects are treated as static and the 4D representation cannot recover them.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised 4D Gaussians best box-labeled splatting","No boxes needed: Neural motion flow drives dynamic splatting","Dynamic urban scenes without tracking via self-supervised 4D Gaussians","Motion flow field turns 4D Gaussians self-supervised on road scenes","SplatFlow: Label-free dynamic splatting outdoes supervised baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1298,"prompt_tokens":967,"completion_tokens":331,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":235}},"tokens_in":583,"tokens_out":331,"duration_ms":3769,"temperature":1.0,"reasoning_tokens":235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:14:16.152580+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a scene containing a small, mostly occluded vehicle moving in 3D but slowly in image space; render it with SplatFlow, then render the same scene after forcing the dynamic mask from ground-truth 3D bounding boxes. If the flow-based mask misses the vehicle, the dynamic-region PSNR will drop sharply relative to the box-forced version, while static metrics stay similar.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time 4 Figure 15","cited_arxiv_id":null,"evidence_quote":"Base 3D Gaussian splatting representation and differentiable rasterizer that the 4D formulation extends."},{"cited_title":"Scene flow from point clouds with or without learning","cited_arxiv_id":null,"evidence_quote":"Supplies pseudo scene-flow labels from point clouds used to pretrain NMFF with bidirectional Chamfer distance."},{"cited_title":"S-nerf: Neural radiance fields for street views","cited_arxiv_id":null,"evidence_quote":"Box-supervised Gaussian baseline; its reliance on tracked 3D bounding boxes is the scalability problem SplatFlow removes."}],"review_version":1}