{"id":"bc43d6a2-e9a6-4f67-8db3-403a5074d300","arxiv_id":"2506.02751","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"RobustSplat improves transient-free 3D Gaussian Splatting by postponing densification to 10,000 iterations and bootstrapping mask supervision from low to high resolution.","lead":"A 3D Gaussian Splatting method for scenes with moving objects delays the point-growing step until the static background is well reconstructed, and trains its transient-object mask at low resolution first, then high resolution. On standard benchmarks it reports higher PSNR and cleaner renderings than SpotLessSplats, WildGaussians, and T-3DGS.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Delayed densification and mask bootstrapping are confounded: the 10K densification start is tuned on the evaluation scenes, so the central claim that delayed growth (not hyperparameter selection) drives the gains is not yet supported.","rationale":"The paper is an empirical systems paper with a clean, well-ablated pipeline, and the reader's conditional verdict is appropriate. My strongest concern overlaps with the reader's weakest assumption (scale-cascaded mask bootstrapping and transferability) but I locate the decisive risk slightly differently: not whether low-resolution DINOv2 features suppress local noise in general, but whether the specific configuration pair (densification start and resolution schedule) is causally responsible for the reported wins or is a tuned artifact. The evidence that would settle this is the same kind the paper already uses (ablations), but evaluated with a selection protocol that separates the tuning surface from the evaluation surface. This concern is about correctness risk, not about consensus disagreement or internal inconsistency. I do not think the paper's claim should be rejected: the mechanism is plausible, the ablations are directionally consistent, delayed densification alone improves 3DGS (Table S3), and the method is faster than several baselines. However, absent code, seeds, or held-out tuning, a CONDITIONAL verdict demanding the selection-protocol check and error bars is the honest disposition. The reader identified hyperparameter tuning on evaluation datasets and missing error bars among their reasons for CONDITIONAL; I agree with those, and my concrete test sharpens them into one decisive experiment.","tokens_in":16329,"tokens_out":1704,"duration_ms":17866,"concrete_test":"Run the full method and ablations exactly as specified, but with the 10K densification start and the 224/504 resolution schedule chosen on a separate validation split (e.g., hold out two of the six On-the-go scenes or use RobustNeRF for selection), then freeze these hyperparameters and report results on all twelve On-the-go scenes plus RobustNeRF. If the reported ranking holds on the untouched scenes and the average PSNR margin over T-3DGS remains similar, the delayed-growth/bootstrapping story survives; if the margin collapses or reverses, the headline result is substantially owed to tuned hyperparameters.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the proposed combination achieves best results on NeRF On-the-go, RobustNeRF, and On-the-go II, and that the two designs are causally responsible. The load-bearing assumption is that the specific 10K densification start and the 224-to-504 resolution schedule transfer across scenes and datasets rather than being fitted to the benchmarks. The paper's own evidence undercuts this: Fig. 5 sweeps densification start (500/2K/5K/10K) and all ablations in Tables 2 and S3 are on the same six On-the-go scenes used for the sweep. With no held-out tuning scenes, no multi-seed variance, and no error bars, the comparison is vulnerable to selection: a configuration tuned on the test set can rank first without the mechanism being the cause. The subsidiary claim that disabling densification is already competitive with SpotLessSplats is itself a single observation (Fig. 2) without quantitative table support, so the causal story 'densification causes transient overfitting, hence delay it' is plausible but not isolated from the mask MLP and the opacity-reset change (delayed to 15K, Supplementary B). Additionally, hyperparameters such as beta_reg=2000, lambda_reg=2.0, tau_u/tau_l, dilation kernel 7, and per-scene downsampling factors (e.g., factor 4 for arcdetriomphe and patio) are fixed only after development on these datasets. Thus, the strongest claim is not yet robustly established: the mechanism may be correct, but the current evidence cannot rule out benchmark-specific tuning as the source of the reported margins.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper identifies the densification stage of 3D Gaussian Splatting as a source of transient-object artifacts and proposes RobustSplat, which (i) postpones Gaussian densification to iteration 10,000 and (ii) supervises a transient-mask MLP with progressively increasing resolution of DINOv2 feature-similarity and residual losses. The method is evaluated on NeRF On-the-go, RobustNeRF, and a harder On-the-go II variant, reporting best average PSNR/SSIM/LPIPS and consistent qualitative improvements.","tokens_in":16641,"tokens_out":5847,"duration_ms":56769,"significance":"If validated, the delayed-growth insight is a simple and transferable design principle for distractor-free 3DGS. The paper includes extensive comparisons against four recent baselines and ablations of each component. However, the current evidence leaves open the possibility that the reported gains come from test-set hyperparameter selection rather than from the delayed-growth mechanism itself; the recommended experiments are feasible within the paper's scope.","major_comments":[{"comment":"The densification start iteration is chosen by sweeping over the same six NeRF On-the-go scenes that are later used as the evaluation benchmark (Table 1). The 10K value is therefore a test-set-tuned hyperparameter, and the statement that 'delayed Gaussian growth' is causally responsible for the gains is not yet supported. Please provide a validation split, leave-one-scene-out analysis, or cross-dataset transfer evidence, and report multi-run statistics with error bars.","section":"Section 3.3, Fig. 5; Section 4.1, Table 1"},{"comment":"The ablation labeled 'Delayed Gaussian Growth' (DG) changes three optimization aspects at once: densification start (500 to 10K), opacity reset start (3000 to 15,000), and pruning start (to 10,000). The reported DG improvement therefore conflates the effect of densification timing with changes in opacity reset and pruning. Please isolate the densification start by holding opacity reset and pruning at original values (or by ablating each change separately).","section":"Section 4.3, Table 2; Supplementary B"},{"comment":"The ablation for mask bootstrapping (MB) removes the coarse-to-fine supervision completely; it does not test the paper's specific claim that lower-resolution feature similarity is more robust and that a progressive schedule is beneficial. A comparison among constant low-resolution, constant high-resolution, and different switch iterations/resolution pairs is needed to substantiate Fig. 6 and the design in Eqs. (4)-(5).","section":"Section 3.4, Table 2"},{"comment":"The motivating observation that disabling densification in vanilla 3DGS 'yields results comparable to SpotLessSplats' is presented for a single scene without quantitative metrics. This claim is the basis for the delayed-growth design; please provide a per-scene table with PSNR/SSIM/LPIPS for 3DGS without densification.","section":"Section 1, Fig. 2"}],"minor_comments":[{"comment":"The word 'trainsient' should be 'transient'.","section":"Section 2"},{"comment":"The phrase 'the right term is 1 if i = 0' is unclear; the term e^{-i/beta_reg} equals 1 at i=0, but the sentence as written could be read as referring to ||1-M_t||.","section":"Eq. (7) and surrounding text"},{"comment":"The runtime table does not include Robust3DGaussians and T-3DGS; please add them or state why they are omitted for parity with the main comparison.","section":"Supplementary Table S1"},{"comment":"The caption does not explain the 'Coarse/Fine' annotation; please clarify how it maps to the scale-cascaded supervision.","section":"Fig. 3 caption"},{"comment":"The upper and lower parts of Table 4 share the same caption; consider separating them for readability.","section":"Section 4.3, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about selection-on-test-set is, in my reading, the most serious issue and needs to be resolved before publication. The paper's experiments are extensive and the method is promising, but the absence of held-out tuning and the confounded DG ablation make the causal claims premature. I would support acceptance after a major revision that adds the recommended validation experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zoe, here's my read on RobustSplat. The thing to know: it's a genuinely simple idea that works in practice. Delaying Gaussian densification until 10K iterations, and switching mask supervision from low-res to high-res DINOv2 features, gives clear gains over SpotLessSplats, WildGaussians, Robust3DGS, and T-3DGS across NeRF On-the-go, RobustNeRF, and On-the-go II. That's a real result, not a marginal one.\n\nWhat's actually new: the densification analysis in Fig 2, showing that vanilla 3DGS with densification disabled already mitigates transient artifacts, is a clean observation that hasn't been made explicitly before. The delayed-growth scheduling is a small but transferable design change. The scale-cascaded mask bootstrapping is a neat reuse of coarse-to-fine training, and the ablations in Tables 2, S2, S3 and 4 show each component contributes. For an empirical paper in this area, the ablation coverage is good.\n\nThe soft spots are real but not disqualifying. The biggest: the densification start (10K) and the resolution schedule (224 to 504) were selected by sweeping on the same six test scenes used in the ablations. Without held-out scenes, multi-seed runs, or error bars, you can't rule out benchmark-specific tuning as part of the margin. The stress-test takes the strong form — the causal story is confounded — and I think that's fair but not fatal. The mechanism (early densification fits transients) is plausible and backed by Fig 5 and Table S3, but the exact magnitude is uncertain. Second, the manual segmentation of test images in On-the-go II and the lack of released code make it harder to verify. Third, the 'comparable to SpotLessSplats' claim for densification-disabled 3DGS is visual only, though Table S3 gives quantitative support for the delayed version. Minor: the limitations section honestly notes illumination changes aren't handled, which is fine.\n\nWho's this for? Anyone working on in-the-wild 3DGS. It deserves a serious referee; the empirical result is strong enough to engage with, even if the causal interpretation needs tightening. I'd ask for multi-seed variance and a held-out tuning split before accepting. But desk-rejecting this would be wrong.","headline":"Delayed densification plus coarse-to-fine mask supervision is a simple, plausible recipe that beats current 3DGS baselines on transient scenes, though test-set-tuned hyperparameters and no error bars mean the causal story isn't nailed down.","tokens_in":17268,"tokens_out":2207,"would_cite":true,"duration_ms":21947,"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":"Postponing Gaussian densification to iteration 10,000, with a transient mask first supervised at low resolution and then at high resolution, yields the best distractor-free 3D Gaussian Splatting reconstructions on the NeRF On-the-go…","keywords":["3D Gaussian Splatting","transient objects","densification scheduling","DINOv2 features","transient mask estimation","novel view synthesis","distractor removal","in-the-wild scenes"],"falsifier":"Hand-label the transient objects in a NeRF On-the-go scene, then measure the mask MLP's IoU against those labels during the first 10,000 iterations under low-resolution versus high-resolution supervision with all other settings fixed; if the low-resolution version does not reduce false positives on under-reconstructed static regions, the scale cascade is not doing the claimed work. A second decisive test is a scene where a transient object stays nearly stationary for most of the training frames: under delayed densification that object may be absorbed into static structure before the mask can filter it, and the final rendering would keep it rather than remove it.","tokens_in":16058,"feed_emoji":"🧹","tokens_out":8968,"duration_ms":78459,"temperature":0.7,"pith_summary":"The paper's central claim is that the adaptive densification built into 3D Gaussian Splatting is a hidden cause of transient-object artifacts: because splitting and cloning starts at iteration 500 in vanilla 3DGS, newly grown Gaussians latch onto moving people, vehicles, and other distractors, and the resulting ghosts persist in novel views. The authors support this diagnosis by showing that running plain 3DGS with densification disabled already suppresses these artifacts about as well as SpotLessSplats, a dedicated robust method. Their proposed fix, RobustSplat, makes two scheduling changes: postpone densification to iteration 10,000 so the static scene structure converges first, and supervise the transient mask MLP with DINOv2 feature cosine similarity at low resolution (224×224) before switching to high resolution (504×504), because low-resolution features tolerate under-reconstructed static regions better. On the NeRF On-the-go benchmark the method claims the best PSNR, SSIM, and LPIPS on all six standard scenes, and the best averages on RobustNeRF and the harder On-the-go II collection. If the claim is right, unstructured photo collections can be rendered cleanly with two schedule tweaks rather than a redesigned network.","feed_headline":"Postponing Gaussian splitting gives cleanest transient-free 3D views","feed_subtitle":"Two schedule tweaks beat prior robust methods on all six NeRF On-the-go scenes.","key_machinery":"Two scheduling mechanisms carry the argument. Delayed Gaussian growth moves the start of 3DGS adaptive density control from iteration 500 to iteration 10,000 (with pruning and opacity reset shifted to match), so early optimization adjusts the color and shape of existing Gaussians without spawning new primitives that could encode transient objects. Scale-cascaded mask bootstrapping is a curriculum for the transient mask MLP, a two-layer network conditioned on DINOv2 features: it is supervised first by image residuals and the cosine similarity between rendered and captured DINOv2 features at 224×224 resolution, then switched to 504×504 once densification starts, exploiting the fact that low-resolution feature patches cover broader context and are less sensitive to local noise. Together they break the negative feedback loop in which an under-reconstructed static scene mislabels its own smooth regions as transient, which then blocks those regions from ever being reconstructed.","core_discovery":"The paper's discovery is a coupling between Gaussian densification and transient-object fitting: densification is what converts moving objects from a mild per-frame inconsistency into persistent 3D artifacts. In vanilla 3DGS, adaptive density control begins at iteration 500, and as training proceeds the newly split and cloned Gaussians increasingly reproduce the transient content in each view, which is why quality degrades over time rather than improving. Disabling densification entirely removes the artifacts and matches a purpose-built robust method, but leaves scenes blurry where the initial SfM point cloud is sparse. RobustSplat therefore keeps densification but delays it to iteration 10,000, and pairs this with a scale-cascaded mask bootstrapping scheme: the mask MLP is first trained on low-resolution image residuals and DINOv2 feature similarity, which integrate broader context and suppress local noise, and then on high-resolution supervision once densification begins, which recovers fine mask boundaries. The paper reports that this combination achieves the best PSNR, SSIM, and LPIPS on all six standard NeRF On-the-go scenes and the best averages on RobustNeRF and NeRF On-the-go II when compared with SpotLessSplats, WildGaussians, Robust3DGaussians, and T-3DGS.","pith_inferences":["The diagnosis generalizes beyond 3DGS: in any adaptive representation where new capacity is allocated based on current reconstruction error, premature capacity allocation can encode outliers before the consensus structure is known, so deferring allocation may be a general robustness lever for neural rendering.","The fixed 224-to-504 switch is a discrete version of a supervision-noise curriculum; a natural extension is to tie the switch to measured reconstruction error or to fade resolution continuously rather than stepping it.","If the proposed mechanism is the true cause, the benefit of delayed growth should scale with the fraction and motion of transient content; on nearly static scenes it should neither help nor hurt much, which is a cheap experiment that would confirm or stress the mechanism.","Because the mask signal is DINOv2 cosine similarity, the method inherits DINOv2's blind spots, such as consistent failures on rare or unusual objects; ensembling two feature sources might close that gap, but the paper only tests alternatives as single replacements."],"forward_implications":["On unstructured photo collections with moving people or vehicles, a robust 3DGS pipeline can be obtained by changing the densification schedule instead of adding a heavier segmentation or uncertainty network: postponing growth and delaying the opacity reset capture most of the gain.","The densification start iteration is shown in the paper's ablations to be a first-order design choice: starts at 10,000 iterations beat starts at 500, 2,000, and 5,000 both with and without mask learning.","Mask supervision resolution acts as a robustness-versus-precision knob, and a coarse-to-fine schedule gives the mask MLP a training curriculum that avoids early misclassification of static regions as transient.","The two schedules attach to the existing feature-based mask paradigm, so they can be used with other feature extractors (the paper tests SAM2 and Stable Diffusion features with a modest drop in performance) rather than being tied to DINOv2 specifically.","The gains come without added per-iteration cost: training runs at roughly 15 minutes per scene on an RTX 3090, faster than the WildGaussians and SpotLessSplats baselines it outperforms."],"supporting_citations":[{"why":"The base 3D Gaussian Splatting representation whose densification schedule is analyzed; the ablation that disables its densification motivates the delayed-growth design.","marker":"[16]"},{"why":"SpotLessSplats, the robust-mask baseline whose image-residual robust loss the mask MLP reuses, and the comparison that plain 3DGS without densification already matches.","marker":"[34]"},{"why":"WildGaussians, which establishes DINO-feature-based uncertainty masks and the cosine-similarity conversion used for the feature robust loss; also a comparison baseline.","marker":"[18]"},{"why":"DINOv2, the feature extractor whose embeddings condition the mask MLP and whose rendered-versus-captured cosine similarity supervises it.","marker":"[28]"},{"why":"The NeRF On-the-go dataset, the primary benchmark including the harder On-the-go II subset.","marker":"[31]"},{"why":"The RobustNeRF dataset used for the second evaluation benchmark.","marker":"[33]"}],"fun_headline_variants":["Delayed densification stops transient ghosts in 3DGS","RobustSplat: delay growth, then sharpen masks for clean 3D views","Splitting later and cascaded masks beat transient artifacts","Two tweaks to 3DGS training yield transient-free rendering","Postponing Gaussian growth cleans up 3D scenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that cosine similarity between DINOv2 features of the rendered and the real image is a reliable and transferable signal for where transient objects are, and specifically that low-resolution features give more robust supervision than high-resolution features during early training; if low-resolution supervision does not actually suppress noise, or the fixed 224-to-504 schedule does not transfer across scenes, the mask will misclassify static regions and the delayed-growth schedule alone will not deliver the reported gains.","fun_headline_variants_meta":{"raw":{"variants":["Delayed densification stops transient ghosts in 3DGS","RobustSplat: delay growth, then sharpen masks for clean 3D views","Splitting later and cascaded masks beat transient artifacts","Two tweaks to 3DGS training yield transient-free rendering","Postponing Gaussian growth cleans up 3D scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000642,"raw_usage":{"total_tokens":2989,"prompt_tokens":1018,"completion_tokens":1971,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":1881}},"tokens_in":634,"tokens_out":1971,"duration_ms":14100,"temperature":1.0,"reasoning_tokens":1881,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:16:37.173364+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hand-label the transient objects in a NeRF On-the-go scene, then measure the mask MLP's IoU against those labels during the first 10,000 iterations under low-resolution versus high-resolution supervision with all other settings fixed; if the low-resolution version does not reduce false positives on under-reconstructed static regions, the scale cascade is not doing the claimed work. A second decisive test is a scene where a transient object stays nearly stationary for most of the training frames: under delayed densification that object may be absorbed into static structure before the mask can filter it, and the final rendering would keep it rather than remove it.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"The base 3D Gaussian Splatting representation whose densification schedule is analyzed; the ablation that disables its densification motivates the delayed-growth design."},{"cited_title":"Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild","cited_arxiv_id":null,"evidence_quote":"The NeRF On-the-go dataset, the primary benchmark including the harder On-the-go II subset."},{"cited_title":"Robustnerf: Ignor- ing distractors with robust losses","cited_arxiv_id":null,"evidence_quote":"The RobustNeRF dataset used for the second evaluation benchmark."}],"review_version":1}