{"id":"d3a64663-9439-4f56-9473-cbc53cc08936","arxiv_id":"2505.13839","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MGStream identifies motion-related Gaussians from a motion mask and convex hull, deforms only those per frame, and optimizes color for emerging objects, improving streaming dynamic view synthesis.","lead":"MGStream separates a 3D scene into static and motion-related Gaussians, deforming only the moving ones for each new video frame. It reports better rendering quality, lower storage, and reduced flicker compared to prior streaming methods on two captured datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The motion mask (Eq. 3) with fixed thresholds and morphological smoothing has unknown recall; any missed motion is never updated, so the central claim of superior quality and temporal consistency is not established for scenes with thin, fast, or low-texture motion.","rationale":"The reader's weakest assumption points to the motion mask recall, and I agree this is the linchpin. The paper's central claim is that restricting deformation/optimization to motion-related Gaussians yields better quality, storage, and temporal consistency. For that to be true, the mask must not miss any motion that would otherwise require updating. The paper provides no analysis of false negatives, nor any robustness study over threshold/kernel choices. My proposed test measures recall against a ground-truth change mask derived from the actual rendering error, which directly quantifies the blind spots. If recall is high, the claim is credible; if not, the superiority could be an artifact of the favorable datasets. The concern is not an attack on the method's elegance; it is a request for evidence on the one assumption that, if wrong, breaks the mechanism. The verdict remains CONDITIONAL: the paper is promising but should either provide the recall analysis or add a fallback that updates high-error regions even when the mask missed them.","tokens_in":14556,"tokens_out":9021,"duration_ms":89397,"concrete_test":"On a subset of 30 frames from N3DV 'Coffee Martini' and 'Cook Spinach', for each frame t compute a ground-truth change mask from a held-out training view (e.g., camera 1) as the pixels where the render of the previous frame's Gaussians differs from the current frame by more than 5/255. Compare this to the motion mask from Eq. 3 using the paper's thresholds (optical flow threshold 1, temporal difference 10, kernel size 20). Report recall (fraction of change-mask pixels inside the motion mask) and precision. If recall falls below 98% on any frame with visible motion, the method has blind spots and the quality claim is at risk; if recall is consistently above 98%, the concern is mitigated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"MGStream's entire advantage rests on the motion mask (Eq. 3) correctly identifying every pixel that belongs to moving or emerging content. The mask is formed by thresholding optical flow at 1 px and temporal difference at 10, then applying a 20-pixel morphological kernel, and intersecting the two cues. The paper reports no recall or stability analysis for this mask. Any moving region that falls below these thresholds, or is eroded by the morphological operations, is never assigned to Gm; consequently it is neither deformed (Eq. 8) nor color-optimized (Eqs. 9-11), and the static Gaussians for that region are never updated after frame 0. The result is that the previous frame's pixels persist, producing ghosting or stale appearance. The problem is structural, not a tuning detail: there is no feedback loop that revisits missed regions, because the optimization loss L_color is only applied to the selected Gm and the static Gaussians are frozen. The two evaluated datasets (N3DV, MeetRoom) contain relatively slow, textured, centrally placed motion; the claim to 'surpass existing streaming 3DGS-based approaches' in general requires evidence that the mask has high recall on harder cases (thin limbs, low-texture objects, fast motion, appearance changes like lighting). Without such evidence, the favorable comparisons may be dataset-specific.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MGStream, a per-frame streaming method for dynamic novel view synthesis with 3D Gaussians. It computes a 2D motion mask by intersecting a thresholded optical-flow map with a morphologically smoothed frame-difference map (Eq. 3), back-projects the mask to Gaussian IDs via alpha-blending ID maps, expands the selected set with a clustering and convex-hull step (Eqs. 4-7), and then deforms only the motion-related Gaussians with a per-frame rigid transformation (Eq. 8). An attention map based on the rendered L1 error selects a subset of those Gaussians whose spherical-harmonic colors are further optimized for emerging content (Eqs. 9-11). Static Gaussians are frozen. Experiments on N3DV and MeetRoom report improved PSNR, lower storage, shorter training time, and lower warping error relative to StreamRF, Dynamic-3DGS, and 3DGStream, with ablations in Tables 3-5.","tokens_in":14831,"tokens_out":5893,"duration_ms":59110,"significance":"If the mechanism is robust, the paper's separation of motion-related and static Gaussians is an elegant and practical idea: it freezes the static backbone, which plausibly explains the reported flicker reduction and storage savings, and it keeps the per-frame training cost low. The paper ships an open-source implementation, includes component ablations, and reports a wide range of metrics. The core risk is that the entire pipeline is gated by one hand-thresholded motion mask whose recall is never measured, and the evaluation is limited to two benign datasets; the central claim of general superiority over streaming 3DGS methods therefore needs additional evidence.","major_comments":[{"comment":"The motion mask is the sole gate for all updates to the dynamic content, and the static Gaussians are never revisited by any loss. The mask is formed by thresholding optical flow at 1 pixel and temporal difference at 10, then applying morphological DILATE and ERODE with a kernel size of 20 and intersecting the two cues. The paper does not report the recall or stability of this mask. Any moving or emerging region that falls below these thresholds, or is eroded by the morphological operations, is neither deformed per Eq. (8) nor color-optimized per Eqs. (9)-(11), so the previous frame's content persists and produces ghosts or stale appearance. Because the two datasets contain relatively slow, textured, centrally placed motion, the favorable comparisons in Table 1 may be dataset-specific. Please provide a recall/precision analysis of the motion mask against a reference, a sensitivity study over the thresholds and kernel size, and either a fallback mechanism that revisits high-error static regions or evidence that such a mechanism is unnecessary.","section":"§4.1, Eq. (3)"},{"comment":"The headline comparisons lack error bars, and the dynamic-region PSNR is stated to be defined in the supplementary material, which is not available in the submitted manuscript. The differences between methods are small (for example, 31.84 versus 32.02 PSNR on N3DV), and the ablations in Tables 3-5 do not state which dataset or scene they use, making it impossible to judge how the conclusions transfer across scenes. Please report results over multiple runs with variance, move the dynamic-region evaluation protocol into the main text, and specify the dataset and scene used for each ablation.","section":"§5.2, Tables 1 and 3-5"},{"comment":"The optimization stage updates only the spherical-harmonic colors of the Gaussians selected by the attention mask; it does not update position, scale, opacity, or density, and no densification mechanism is mentioned. For an object that emerges in a region not already covered by motion-related Gaussians from the previous frame, color-only updates can change the appearance but cannot create new geometry. The ablation in Table 4 shows that adding color optimization improves PSNR from 33.51 to 34.31, but this does not establish that the method handles genuinely new geometry. Please clarify whether emergence is always confined to regions already represented by Gm, or provide experiments with an object entering the field of view from outside.","section":"§4.2, Eqs. (9)-(11)"}],"minor_comments":[{"comment":"The term 'wrap-based methods' appears twice in Section 2.2 and should be 'warp-based methods' to match the terminology used in the Introduction.","section":"§2.2"},{"comment":"The caption contains a typo: 'motio-related 3DGs' should be 'motion-related 3DGs'.","section":"Fig. 4 caption"},{"comment":"The phrase 'The detail abouts PSNR calculation' should be 'The details of the PSNR calculation'.","section":"Table 1 caption"},{"comment":"The symbols D and E for DILATE and ERODE are used without defining the structuring element or kernel shape; please state these explicitly.","section":"Eq. (3)"},{"comment":"The ablation tables would be easier to interpret if they stated the dataset, scene, and the number of frames averaged, and if the units of Ewarp were given.","section":"Tables 3-5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript repeatedly refers to a supplementary material section for the dynamic-region PSNR protocol and additional visualizations, but no supplement is present in the submitted version. I would ask the editor to verify that the supplement is included in any resubmission, since Table 1's secondary metric depends on it. The novelty of the static/dynamic Gaussian decomposition itself is clear, and I do not see a disclosure concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you're following streamable dynamic NVS. The core idea is simple and effective: instead of deforming all Gaussians each frame like 3DGStream, MGStream builds a motion mask from optical flow and frame difference, back-projects it to Gaussians via Gaussian ID maps, then uses DBSCAN clustering plus a Delaunay convex hull to include interior Gaussians. Only those motion-related Gaussians get a rigid deformation and, for emerging objects, an attention-gated SH color optimization. Static Gaussians stay frozen. That design delivers the main result: per-frame storage drops from 4.0 to 0.7 MB on MeetRoom and 7.6 to 2.1 MB on N3DV relative to 3DGStream, with slightly better PSNR and lower Ewarp. The ablations support each component; the convex hull and clustering are what let the method cover full moving objects rather than just surfaces.\n\nWhat's genuinely new is the combination—GIM back-projection plus clustering-based convex hull for identifying motion-related Gaussians, and the attention-based SH update for emerging content. That is a concrete, reproducible mechanism, and code is provided. It is not a rehash of prior work.\n\nThe soft spots are real but not fatal. The biggest is the motion mask in Eq. 3: fixed thresholds (flow at 1px, difference at 10) plus a 20-pixel morphological kernel, then an intersection. There is no recall or stability analysis. Any missed motion is never updated, since static Gaussians are frozen and the loss only touches selected Gm. On the two datasets—slow, textured, centrally located people—the mask works well, but the paper's general claim is not established for thin, fast, or low-texture motion. A revision should show threshold sensitivity or at least a failure analysis.\n\nSmaller issues: the dynamic-region PSNR is only defined in the supplementary (which I could not check), there are no error bars, the Ewarp improvement over 3DGStream is modest, and the emerging-object optimization only updates color, not geometry, which is a deliberate but under-discussed trade-off. Citation pattern is fine; related work is adequate and concurrent works are mentioned. The motion mask comes from input cues, not the model's output, so there is no circularity in the main evaluation, and the attention mask is an internal training signal, not a metric.\n\nBottom line: this is a solid, honest paper that deserves a serious referee. The central claim holds up on the tested datasets, with limitations a revision should address. I'd send it to review and would bring it to our reading group.","headline":"MGStream is a credible incremental step for streamable dynamic NVS: motion-mask-guided deformation cuts storage 3-4x with equal or better quality, but the motion mask's recall is the main limitation to probe in review.","tokens_in":15373,"tokens_out":4163,"would_cite":true,"duration_ms":36405,"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":"MGStream claims that streaming dynamic scenes can be compressed and stabilized by updating only the Gaussians that move.","keywords":["dynamic novel view synthesis","3D Gaussian Splatting","streamable reconstruction","motion mask","per-frame training","temporal consistency","emerging objects","storage efficiency"],"falsifier":"Render a dynamic sequence containing a thin, fast-moving object (such as a swinging cord) or a slow color change that falls below the optical-flow and frame-difference thresholds, and inspect the frames for stale background or ghosting exactly at the masked-out pixels; if such artifacts appear, the central claim that motion-related Gaussians fully cover the dynamics is false. More directly, compare the motion mask against hand-labeled moving-object masks on the evaluation views and measure its recall.","tokens_in":14355,"feed_emoji":"🎥","tokens_out":7281,"duration_ms":64219,"temperature":0.7,"pith_summary":"MGStream's central claim is that streaming dynamic-scene reconstruction from multi-view video can be made cheaper and temporally stabler by updating only the 3D Gaussians that correspond to motion, instead of deforming and re-optimizing the entire scene every frame. Given two consecutive frames, it builds a motion mask from optical flow and temporal differences, maps that mask back onto the Gaussian set, and expands the set with a clustering-based convex hull so Gaussians inside moving objects are included. Only those motion-related Gaussians are rigidly deformed, and only a further attention-selected subset has its color coefficients optimized for newly appearing content; static Gaussians are left untouched. The motivated payoff is practical: online per-frame training with less stored data, fewer flicker artifacts, and quality at least matching methods that update everything. Experiments on two real-world multi-camera datasets support these claims on image quality, warping error, training time, and storage.","feed_headline":"Moving-only Gaussian updates cut flicker and storage in streaming","feed_subtitle":"MGStream beats per-frame streaming baselines on quality, temporal consistency, and stored size.","key_machinery":"The load-bearing object is the motion-related Gaussian set $G_m$, selected through a four-step chain: a motion mask from optical flow and temporal difference (Eq. (3)); Gaussian ID maps that record which Gaussian dominates each pixel's $\\alpha$ blending (Eq. (5)); density-based clustering followed by Delaunay-based convex hulls to pull in the Gaussians inside the moving volume (Eqs. (6)-(7)); and, for emerging objects, an attention mask computed from large per-pixel errors after deformation (Eq. (9)) that selects the subset $G_{\\text{new}}$ for spherical harmonic color updates (Eqs. (10)-(11)). This chain does the work of confining all per-frame deformation and optimization to a small, scene-relevant fraction of the representation, which is the source of both the storage savings and the temporal consistency.","core_discovery":"MGStream is built on the idea that a dynamic scene can be split into a frozen static Gaussian model and a small per-frame set of motion-related Gaussians. To locate that set, the paper computes a motion mask $\\hat{M}$ (Eq. (3)) by thresholding optical flow and frame difference, then applying morphological dilation and erosion; Gaussian ID maps (Eq. (5)) back-project the masked pixels to the surface Gaussians $G_o$; and density-based clustering plus Delaunay-based convex hulls (Eqs. (6)-(7)) add the interior Gaussians $G_i$, yielding $G_m$. The dynamic is then modeled by rigid translation and rotation offsets from a hash-grid deformation field applied only to $G_m$ (Eq. (8)), and emerging objects are handled by an attention map (Eq. (9)) that selects a subset $G_{\\text{new}}$ whose spherical harmonic coefficients are updated (Eqs. (10)-(11)). Because the static Gaussians are never touched, the rendered background stays consistent across frames, which is the mechanism behind the reported flicker reduction and storage savings.","pith_inferences":["A natural follow-up is to make the motion mask adaptive: learned or per-scene thresholds would likely improve recall for slow and thin motions, which fixed thresholds can miss.","The same 'update only what moves' separation could be transferred to other primitive-based streaming representations by replacing Gaussian ID maps with per-primitive visibility or hit counts from the rasterizer.","A direct stress test would measure motion-mask recall against labeled object masks on the evaluation datasets; dynamic-region PSNR should track that recall.","The reported storage reductions suggest that quantizing or entropy-coding the per-frame deltas could lower the streaming bitrate further."],"forward_implications":["Per-frame storage scales with the amount of motion in the scene rather than with scene size, since only the selected motion-related Gaussians receive new offsets and color updates.","Static regions of the rendered video remain unchanged across frames, which is what lowers flicker metrics such as Ewarp.","Emerging objects can be rendered without adding or densifying Gaussians, by updating the spherical harmonic colors of already-selected motion Gaussians under an attention mask.","The pipeline remains causal and online: each frame is processed from the previous frame's state without access to future frames.","Training time stays close to the fastest streaming baselines because the deformation and optimization stages run for only 100 iterations each on small Gaussian subsets."],"supporting_citations":[{"why":"Supplies the 3D Gaussian scene representation, differentiable rasterizer, and photometric loss that MGStream builds on and initializes from.","marker":"[29]"},{"why":"The per-frame streaming 3DGS baseline that deforms all Gaussians; MGStream's main comparison and the source of its rigid-deformation design.","marker":"[54]"},{"why":"Streaming dynamic-3DGS baseline whose deformation-on-all-Gaussians approach causes the flicker MGStream targets, and a comparison point.","marker":"[41]"},{"why":"Introduces per-frame streaming radiance fields and the MeetRoom dataset; its method and metrics form a baseline.","marker":"[33]"},{"why":"Supplies the N3DV multi-view video dataset used for evaluation and comparison with prior work.","marker":"[34]"},{"why":"Provides the pre-trained optical-flow model used to compute the flow component of the motion mask.","marker":"[62]"},{"why":"Density-based clustering used in the clustering-based convex hull step to group the back-projected motion Gaussians.","marker":"[17]"},{"why":"Delaunay triangulation implementation used to build the convex structure that captures interior motion-related Gaussians.","marker":"[7]"}],"fun_headline_variants":["Static-motion Gaussian split kills flicker in streaming","Moving-only Gaussians fix flicker in dynamic stream rendering","Streamable 3DGS with motion-aware split trims flicker","MGStream: static-motion split for stable streaming novel views","Flicker-free streaming from motion-aware Gaussian decomposition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the motion mask produced by fixed thresholds on optical flow and frame difference, after morphological dilation and erosion, marks every pixel belonging to a moving or newly appearing object, so that no pixel requiring an update is left to the static Gaussians.","fun_headline_variants_meta":{"raw":{"variants":["Static-motion Gaussian split kills flicker in streaming","Moving-only Gaussians fix flicker in dynamic stream rendering","Streamable 3DGS with motion-aware split trims flicker","MGStream: static-motion split for stable streaming novel views","Flicker-free streaming from motion-aware Gaussian decomposition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001259,"raw_usage":{"total_tokens":5186,"prompt_tokens":1001,"completion_tokens":4185,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":4102}},"tokens_in":617,"tokens_out":4185,"duration_ms":23008,"temperature":1.0,"reasoning_tokens":4102,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:08:59.460606+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a dynamic sequence containing a thin, fast-moving object (such as a swinging cord) or a slow color change that falls below the optical-flow and frame-difference thresholds, and inspect the frames for stale background or ghosting exactly at the masked-out pixels; if such artifacts appear, the central claim that motion-related Gaussians fully cover the dynamics is false. More directly, compare the motion mask against hand-labeled moving-object masks on the evaluation views and measure its recall.","supporting_citations":[{"cited_title":"3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free- viewpoint videos","cited_arxiv_id":null,"evidence_quote":"The per-frame streaming 3DGS baseline that deforms all Gaussians; MGStream's main comparison and the source of its rigid-deformation design."},{"cited_title":"Streaming radiance fields for 3d video synthe- sis","cited_arxiv_id":null,"evidence_quote":"Introduces per-frame streaming radiance fields and the MeetRoom dataset; its method and metrics form a baseline."},{"cited_title":"Neural 3d video synthesis from multi-view video","cited_arxiv_id":null,"evidence_quote":"Supplies the N3DV multi-view video dataset used for evaluation and comparison with prior work."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Delaunay triangulation implementation used to build the convex structure that captures interior motion-related Gaussians."}],"review_version":1}