{"id":"6c491364-a4ef-451d-b021-fe1ef66f39fb","arxiv_id":"2508.14014","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"During online Gaussian splatting SLAM, training extra on non-keyframes that view the most uncertain Gaussians improves model completeness over keyframe-only training.","lead":"A system that builds 3D models from live video now chooses its own training views, picking the frames that look at the least complete parts of the scene. The result is a fuller 3D model from a single camera, aimed at robots, phones, and augmented reality.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The completeness gain is not established as geometric: NVS adds color-only training on non-keyframes, and the paper's own Table S5 shows depth L1 degrades when view count rises; no multi-scene geometric completeness metric is reported.","rationale":"The paper describes a coherent system: an online MVS-3DGS SLAM with independent frontend/backend, GBA, Gaussian deformation, disparity-map integration, and honest ablation tables. The weakest load-bearing point is not the uncertainty metric itself but the leap from 'rendering quality on selected views' to 'enhancing completeness' of the 3D model. Completeness is a geometric property, and the only geometric number reported with NVS ablations is depth L1 from Table S5, which deteriorates when the NVS count goes from 100 to 150 and is not reported as a direct NVS on/off comparison. Table S8 (one scene) is the only direct NVS-vs-random/quality ablation, and it lacks variance and multi-scene support. The text explicitly states that absence of depth in novel views reduces reconstruction quality, so the 100-view choice is a tuned trade-off. A multi-scene geometric completeness evaluation would settle this. Since this is a concern about evidence strength rather than a demonstrated internal contradiction, the reader's CONDITIONAL verdict stands; I do not move it.","tokens_in":9224,"tokens_out":3914,"duration_ms":41061,"concrete_test":"Re-run the NVS on/off comparison on at least the three Replica scenes and two Tanks&Temples scenes used in the paper, and report geometric completeness (e.g., F-score at 1cm/2cm, or accuracy/completeness of the reconstructed depth/point cloud against ground truth) instead of only Depth L1 and rendering metrics. If F-score/completeness does not improve when the 100 uncertainty-selected views are added, the 'enhancing completeness' claim fails and the method reduces to extra color fitting on arbitrary frames.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that adding ~100 uncertainty-selected non-keyframes to keyframe training 'significantly enhances completeness.' What is actually reported are rendering metrics (PSNR/SSIM/LPIPS) and depth L1, plus one single-scene ablation (Table S8, Replica office0). The mechanism by which completeness should improve is not backed by a geometric completeness metric (e.g., F-score, accuracy/completeness on the reconstructed mesh/point cloud). Table S5 shows the cost of the mechanism: novel views are trained with color loss only, and the text says 'the absence of depth information in novel views resulted in a reduction in reconstruction quality.' At 150 novel views, keyframe Depth L1 becomes worse than the 0-view baseline (0.058 vs 0.056); at every-5-frame evaluation it never improves (0.058 vs 0.059). Thus the 100-view operating point is a tuned trade-off, and the claimed completeness improvement could be mostly photometric overfitting to selected views rather than improved 3D geometry. Fig S4 correlates high uncertainty with lower PSNR on one Tanks&Temples scene, but that only shows the selection signal marks hard-to-render views; it does not show that training on those views improves geometric completeness. Without a direct geometric completeness evaluation and multi-scene NVS ablations, the load-bearing premise remains unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an online 3D Gaussian Splatting (3DGS) mapping system for RGB-only image streams. The main novelty is a novel-view selection (NVS) module that estimates per-Gaussian uncertainty from shape and position-gradient statistics, scores incoming non-keyframes by predicted information gain, and adds the highest-uncertainty non-keyframes to the keyframe-based training set. The pipeline combines an MVS frontend (MVSFormer with DROID-SLAM disparity injection) and a parallel 3DGS backend with global bundle adjustment and Gaussian deformation. The paper reports evaluations on Replica, TUM-RGBD, ScanNet, Tanks & Temples, and an aerial dataset, with ablations on the number of selected views, selection criteria, disparity-map use, refinement iterations, and runtime.","tokens_in":9516,"tokens_out":2512,"duration_ms":28161,"significance":"If the central claim were fully established, the contribution would be useful: an online 3DGS mapper that adaptively chooses a small set of informative non-keyframes, improving reconstruction completeness while remaining real-time. The paper includes directed ablations that support the mechanism in a limited setting: uncertainty-based selection beats random and PSNR-based selection on Replica office0 (Table S8, PSNR 43.93 vs 43.29/43.35), disparity-map injection improves both rendering and depth (Table S7), and the system is evaluated on external benchmarks including outdoor scenes. However, the paper's headline claim is about geometric completeness, and that claim is not directly measured. The evidence actually reported is photometric quality and depth L1 on one ablation scene, plus a qualitative TartanAir figure. The runtime prose in §B.9 is also inconsistent with the table it cites. The core idea is plausible, but the current manuscript does not yet provide the geometric and multi-scene support needed to substantiate the completeness claim.","major_comments":[{"comment":"The central claim is that NVS 'significantly enhances completeness,' but no geometric completeness metric is reported anywhere. Table S8, the only quantitative NVS ablation, reports PSNR/SSIM/LPIPS and depth L1 on a single Replica scene (office0); these are rendering and depth accuracy metrics, not completeness (e.g., F-score, accuracy/completeness on a reconstructed mesh/point cloud, or coverage of visible surfaces). Fig. S4 shows that high-uncertainty views have lower PSNR, which supports that the selector identifies hard-to-render viewpoints, but it does not show that training on those views improves geometric completeness. Please add a multi-scene evaluation with direct geometric completeness metrics, or revise the claim to 'improved rendering quality' rather than 'completeness.'","section":"§1, §B.5, Table S8, Fig. S4"},{"comment":"Table S5 is in tension with the mechanism claimed. At the operating point of 100 novel views, keyframe depth L1 improves (0.056→0.050) but every-5-frame depth L1 worsens (0.058→0.059); at 150 views, keyframe depth L1 also worsens (0.058 vs 0.056 baseline). The text states that 'the absence of depth information in novel views resulted in a reduction in reconstruction quality.' Since novel views are trained with color loss only, the evidence supports a trade-off rather than a geometric completeness gain. The manuscript should either provide depth-supervised novel-view training or a direct geometric evaluation that separates photometric overfitting from genuine 3D completion. Without this, the 100-view choice reads as a tuned compromise, not a validated mechanism.","section":"§B.4, Table S5"},{"comment":"The prose claims that the method 'delivers a significantly higher frame rate than MonoGS and Splat-SLAM,' but Table S9 lists Avg. FPS of 10.77 for Ours versus 17.57 for Splat-SLAM. Under the natural reading of the table, Splat-SLAM is about 1.6× faster. Please correct the claim or explain the measurement protocol (e.g., whether the comparison is on a different hardware/config); as written, the text contradicts the table.","section":"§B.9, Table S9"}],"minor_comments":[{"comment":"The table caption does not define '0' in the 'Number of novel views' column; clarify that this is the keyframe-only baseline. Also, the FPS row shows a significant drop (12.07→9.18→8.54) that is discussed in prose but should be more prominent given the real-time claim.","section":"§B.4, Table S5"},{"comment":"The ablation is on a single scene and reports no variance or statistical significance. Please state whether the difference (e.g., depth L1 0.034 vs 0.042) is stable across runs or scenes; otherwise the comparison is anecdotal.","section":"§B.8, Table S8"},{"comment":"The column header 'Avg. FPS↑[GIB]↓' mixes FPS and GPU memory units; split into separate columns for readability. Also 'Table. S9' should be 'Table S9'.","section":"§B.9, Table S9"},{"comment":"The caption uses 'Scannet' while the text and references use 'ScanNet'; make capitalization consistent.","section":"Figure S5"},{"comment":"The reference [Redmon, 2016] appears in the bibliography but is not cited in the visible text; check whether it is needed or remove it.","section":"References"},{"comment":"The introduction states that previous methods store depth only for keyframes, but the contribution over MVS-GS is not crisply stated. Consider a short paragraph explicitly listing the delta from [Lee et al., 2024] (GBA with Gaussian deformation, NVS, disparity fusion) to help readers position the work.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"The paper builds directly on the authors' own MVS-GS work, and the related-work section frames the contribution as an extension; this is legitimate, but the novelty should be assessed carefully relative to MVS-GS and to prior view-selection ideas (e.g., FisherRF, GS-Planner). The NVS ablation is single-scene, and the completeness claim is not backed by geometric evaluation; these are fixable with additional experiments within the paper's scope, hence major revision rather than reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the twist here is real and worth a look. Previous active-view-selection works (FisherRF, GS-Planner, NeurAR) plan camera motion; this paper instead picks already-captured non-keyframes for extra training, using Gaussian shape/position-gradient uncertainty. That is a legitimate extension, not just a repackaging. The system is coherent, and the authors do honest ablations against random and PSNR-based selection, with the uncertainty signal correlating with hard-to-render views (Fig. S4). They also report directed ablations for disparity fusion (Table S7) and final-iteration count (Table S6). No fatal internal contradiction.\n\nWhere it gets soft is the central claim: that NVS \"significantly enhances completeness.\" The evidence is mostly rendering metrics plus one single-scene ablation (Table S8, Replica office0). There is no geometric completeness metric—no F-score or accuracy/completeness on the reconstructed geometry. Table S5 shows the mechanism's cost: novel views are trained with color loss only, and depth L1 degrades when you go from 100 to 150 views. At the 100-view operating point, every-5-frame depth L1 is flat (0.058 vs 0.059), and keyframe depth L1 only improves from 0.056 to 0.050. So the completeness gain could be photometric overfitting to selected views rather than improved 3D geometry. The paper itself says \"the absence of depth information in novel views resulted in a reduction in reconstruction quality.\" That is an honest admission, but it undercuts the main claim.\n\nThere is also a clear internal inconsistency in efficiency reporting. Section B.9 says the method delivers \"a significantly higher frame rate than MonoGS and Splat-SLAM,\" but Table S9 shows 16.34 FPS vs Splat-SLAM's 17.57, and GPU memory usage 10.77 GB vs 1.24 GB. That claim does not survive contact with its own table. And there is no direct comparison to the authors' own MVS-GS base, so it is hard to attribute improvements to NVS versus the other framework changes.\n\nA few lesser points: operating points (100 views, 2K iterations) appear tuned on evaluation scenes, and the single NVS ablation lacks variance. The provided text omits the main method and experiments, so I cannot check the full evaluation, but the appendix carries enough evidence to gauge the thrust.\n\nOverall: the idea has merit and the work is honest about trade-offs, but the load-bearing completeness claim is not yet supported. I would send it to peer review, because the concept is timely and a careful referee could demand the missing geometric evaluations. A revision with multi-scene NVS ablations and a direct geometric completeness metric could make this a solid contribution.","headline":"A genuinely new frame-selection idea for online 3DGS, but the completeness claim is running ahead of the evidence.","tokens_in":10079,"tokens_out":1954,"would_cite":true,"duration_ms":22830,"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":"Selecting non-keyframes that observe the most uncertain Gaussians and training them alongside keyframes makes online 3D Gaussian Splatting maps more complete and higher-quality than keyframe-only training, at real-time rates.","keywords":["3D Gaussian Splatting","online mapping","novel view selection","Gaussian uncertainty","multi-view stereo","RGB-only SLAM","real-time reconstruction","view synthesis"],"falsifier":"Run the same 100-view selection ablation on a low-texture indoor scene and a complex outdoor scene with ground-truth geometry, measuring depth L1 and mesh completeness for uncertainty-selected views versus the same number of randomly selected views; if the uncertainty-selected model does not produce consistently better geometry, the uncertainty signal is tracking MVS noise or texturelessness instead of fixable incompleteness.","tokens_in":9072,"feed_emoji":"🗺️","tokens_out":6932,"duration_ms":64021,"temperature":0.7,"pith_summary":"This paper tries to show that an online 3D Gaussian Splatting (3DGS) model built from RGB-only video can be made more complete by adaptively adding a small number of carefully chosen non-keyframes to the training set. The proposed selection signal is per-Gaussian uncertainty, computed from the shape and position-gradient statistics of each Gaussian, and views that project many high-uncertainty Gaussians are judged to carry the most information gain. Combined with a multi-view-stereo frontend that supplies consistent depth and global bundle adjustment with Gaussian deformation, the method reports better rendering and geometry than keyframe-only training, at real-time frame rates, on indoor benchmarks and on complex outdoor scenes.","feed_headline":"100 uncertain views finish incomplete 3D Gaussian maps","feed_subtitle":"Adding the non-keyframes that see the least-settled Gaussians keeps online 3D mapping real-time and more complete.","key_machinery":"The load-bearing object is the Gaussian uncertainty map: each 3D Gaussian primitive receives an uncertainty value derived from its shape (large or elongated primitives are less trustworthy) and from the magnitude of its position gradients during optimization (still-moving primitives are not yet settled). A candidate non-keyframe's information gain is computed by projecting these uncertain Gaussians into the view and aggregating their uncertainty, so the selector can pick frames that observe the least-reconstructed regions. The companion machinery is the online MVS frontend, which estimates dense depth for keyframes and injects a disparity map into the MVS network, keeping the Gaussian traini","core_discovery":"On the paper's own terms, the discovery is that reconstruction quality can be assessed online from the Gaussian field itself: Gaussians that are large or whose positions are still moving during optimization are marked uncertain, and a view's value is measured by how many of these uncertain Gaussians it observes. Selecting about 100 such non-keyframes and training them together with keyframes refines under-reconstructed regions from new viewpoints, yielding more complete models and higher rendering quality than training on keyframes alone. The paper further claims that integrating a learned multi-view stereo network with a disparity prior, plus continuous global bundle adjustment and Gaussian","pith_inferences":["The uncertainty signal is really a next-best-view score in disguise: the same per-Gaussian uncertainty could steer an autonomous camera's motion toward unexplored regions, turning the offline frame-selection rule into an active reconstruction planner.","Because selected non-keyframes are trained with color only, the observed depth-L1 degradation at 150 added views is likely color-fitting winning over geometry; supplying depth for those views would be a direct test of whether completeness gains can be pushed further without sacrificing geometry.","The selection rule is agnostic to the depth source, so it should transfer to other MVS or depth-prediction backends; a straightforward check is running the same NVS logic on a system with a different frontend and comparing completeness.","Persistent high-uncertainty regions across many selected frames could double as a drift or gap detector for the SLAM system itself, flagging places where the trajectory or the map needs revisiting."],"forward_implications":["Training on roughly 100 non-keyframes chosen by Gaussian uncertainty, together with keyframes, yields better rendering PSNR/SSIM/LPIPS and better depth L1 than keyframe-only training in the experiments reported.","Selected high-uncertainty views tend to be under-reconstructed ones: their rendered PSNR is markedly lower (about 19.6 versus 28.8) than low-uncertainty views, and their uncertainty values are roughly three times higher.","The pipeline stays real-time: with novel-view selection it averages about 9 FPS on the Replica office0 sequence, and every module's runtime stays below its invocation interval.","Injecting the DROID-SLAM disparity map into the MVS network improves both depth L1 (0.067 to 0.050) and rendering quality, so the depth prior is a measurable part of the gain.","Only 2K final-refinement iterations suffice to beat Splat-SLAM's 26K iterations on rendering quality in the reported comparison, indicating the online-trained field starts from a much better initialization."],"supporting_citations":[{"why":"Defines the 3D Gaussian Splatting representation and rasterization that the whole online mapping pipeline optimizes.","marker":"[Kerbl et al., 2023]"},{"why":"MVS-GS is the prior online MVS-based 3DGS mapping framework this work builds on and extends with view selection and global bundle adjustment.","marker":"[Lee et al., 2024]"},{"why":"MVSFormer supplies the multi-view stereo depth estimates used to initialize and update the Gaussians.","marker":"[Cao et al., 2022]"},{"why":"Provides the online multiview stereo depth estimation approach and the Aerial outdoor dataset used for evaluation.","marker":"[Song et al., 2021]"},{"why":"DROID-SLAM supplies frontend tracking and a disparity map that is injected into the MVS network.","marker":"[Teed and Deng, 2021]"},{"why":"FisherRF is the alternative Fisher-information uncertainty metric the paper contrasts its shape/gradient uncertainty with.","marker":"[Jiang et al., 2025]"},{"why":"GS-Planner's PSNR-quality-based view selection is the baseline compared against in the novel-view-selection ablation.","marker":"[Jin et al., 2024]"},{"why":"Replica is the indoor benchmark where the main rendering and depth evaluations and ablations are run.","marker":"[Straub et al., 2019]"},{"why":"Tanks and Temples provides the complex outdoor scene benchmark used to show generalization.","marker":"[Knapitsch et al., 2017]"}],"fun_headline_variants":["Pick 100 views to complete 3D maps in real time","Online 3D mapping: add uncertain-view frames for completeness","Gaussians reveal gaps: select non-keyframes to fill 3D scenes","Unsettled Gaussians guide view selection for better 3D models"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The argument rests on the premise that high Gaussian uncertainty marks places where extra color-only training can actually reduce model error, rather than simply marking textureless surfaces, unreliable stereo depths, or inherently hard-to-render viewpoints.","fun_headline_variants_meta":{"raw":{"variants":["Pick 100 views to complete 3D maps in real time","Online 3D mapping: add uncertain-view frames for completeness","Gaussians reveal gaps: select non-keyframes to fill 3D scenes","Unsettled Gaussians guide view selection for better 3D models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1193,"prompt_tokens":714,"completion_tokens":479,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":401}},"tokens_in":458,"tokens_out":479,"duration_ms":4699,"temperature":1.0,"reasoning_tokens":401,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:46:58.054267+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 100-view selection ablation on a low-texture indoor scene and a complex outdoor scene with ground-truth geometry, measuring depth L1 and mesh completeness for uncertainty-selected views versus the same number of randomly selected views; if the uncertainty-selected model does not produce consistently better geometry, the uncertainty signal is tracking MVS noise or texturelessness instead of fixable incompleteness.","supporting_citations":[{"cited_title":"3d gaus- sian splatting for real-time radiance field rendering.ACM Trans","cited_arxiv_id":null,"evidence_quote":"Defines the 3D Gaussian Splatting representation and rasterization that the whole online mapping pipeline optimizes."},{"cited_title":"View path planning via online multiview stereo for 3-d modeling of large-scale structures.IEEE Transactions on Robotics, 38(1):372–390,","cited_arxiv_id":null,"evidence_quote":"Provides the online multiview stereo depth estimation approach and the Aerial outdoor dataset used for evaluation."},{"cited_title":"Droid- slam: Deep visual slam for monocular, stereo, and rgb-d cameras.Advances in neural information processing sys- tems, 34:16558–16569,","cited_arxiv_id":null,"evidence_quote":"DROID-SLAM supplies frontend tracking and a disparity map that is injected into the MVS network."},{"cited_title":"Fisherrf: Active view selection and mapping with radiance fields using fisher information","cited_arxiv_id":null,"evidence_quote":"FisherRF is the alternative Fisher-information uncertainty metric the paper contrasts its shape/gradient uncertainty with."},{"cited_title":"Gs-planner: A gaussian-splatting-based planning framework for ac- tive high-fidelity reconstruction","cited_arxiv_id":null,"evidence_quote":"GS-Planner's PSNR-quality-based view selection is the baseline compared against in the novel-view-selection ablation."},{"cited_title":"Tanks and temples: Bench- marking large-scale scene reconstruction.ACM Transac- tions on Graphics (ToG), 36(4):1–13,","cited_arxiv_id":null,"evidence_quote":"Tanks and Temples provides the complex outdoor scene benchmark used to show generalization."}],"review_version":1}