{"id":"182eba7f-a2ad-484c-8a42-a44a9765c50a","arxiv_id":"2412.00242","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Uni-SLAM combines a model-free predictive uncertainty, based on volume rendering termination probability, with decoupled hash grids and strategic bundle adjustment to achieve state-of-the-art dense RGB-D SLAM on Replica, ScanNet, and TUM RGB-D.","lead":"Uni-SLAM builds a dense 3D map and tracks a camera in real time using two neural hash grids, one for shape and one for color, and a new uncertainty score that tells the system which pixels to trust. The authors report it beats prior neural SLAM systems on standard indoor benchmarks, with a 25% lower depth error and a higher completion rate on the Replica dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncertainty measure Eq. (5) is not validated under pose error; if β_m stays low for confidently wrong renderings, the outlier-filtering claim and tracking gains are unsupported.","rationale":"The reader's weakest assumption identifies the same mechanism I find most load-bearing: the predictive uncertainty β_m is trusted to indicate pixel data quality, but it is a rendering-termination heuristic, not a calibrated uncertainty, and no experiment demonstrates that it remains reliable when the model is confidently wrong. If the uncertainty filter fails under pose error, then the reweighted tracking loss in Eq. (14) can up-weight exactly the pixels that would entrench drift, undermining the claimed tracking improvements and the robustness motivation of the paper. The benchmark results are consistent and the ablations support the design choices, so I do not reject the paper; the concern is addressable by the proposed controlled experiment. Other issues, such as the disclosed ground-truth first pose, the overstated real-time claim on TUM (2.7 FPS), and the absence of code, are real but secondary: they affect presentation and reproducibility more than the conceptual validity of the uncertainty mechanism. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":28472,"tokens_out":5599,"duration_ms":60564,"concrete_test":"Run a controlled experiment on Replica room0 or TUM fr1/desk: fix the trained model at a mid-trajectory frame, take the current estimated pose, add controlled perturbations (e.g., 1-5 cm translation, 0.5-2 deg rotation), render depth, and compute per-pixel depth error against the ground-truth depth. Then compute the Spearman correlation between β_m (Eq. 5) and per-pixel depth error, and the precision/recall of CF_m = 1 as an outlier filter at the reported β_unc_m. If correlation is not significantly positive or if high-error pixels are not preferentially masked under pose error, the load-bearing assumption fails exactly in the regime the method claims to handle; report whether tracking ATE still improves after removing the uncertainty reweighting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that uncertainty-guided reweighting yields state-of-the-art tracking and mapping depends on the assumption in Sec. 3.2, Eq. (5), that β_m = (1-p(r))^2 is high when a pixel's observation is unreliable and low when it is reliable. The paper validates this only qualitatively (Fig. 3) for invalid-depth pixels and by benchmark-level ablations. It never tests the failure mode in which the model is confidently wrong: after a pose error, a ray sampled from the estimated depth can terminate in a previously learned surface with high accumulated termination probability, giving low β_m and CF_m = 1 (Eq. 7), so the erroneous observation is up-weighted during tracking. Because tracking and mapping are jointly optimized, this can amplify drift instead of filtering outliers. The per-dataset thresholds β_unc_m = 1e-2 and β_unc = 1e-3 (or 2e-3 on TUM) make the behavior dataset-tuned, so benchmark gains do not by themselves establish that the uncertainty measure tracks data quality under the conditions it is meant to handle.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Uni-SLAM, a real-time dense neural implicit RGB-D SLAM system with three main components: (i) a decoupled representation with separate multiresolution hash grids for geometry (SDF) and appearance, both decoded by tiny MLPs and rendered by standard volume rendering; (ii) a model-free \"predictive uncertainty\" defined per pixel as beta_m = (1 - p(r))^2, where p(r) is the accumulated termination probability along the depth-guided sampling ray, used to hard-mask and reweight the SDF, depth, and (during tracking only) color loss terms; and (iii) an uncertainty-triggered mapping schedule combining fixed-frequency global bundle adjustment, image-level-uncertainty-activated local BA, and a co-visibility-based local loop closure optimization (LLCO). The system is evaluated on Replica, ScanNet, and TUM RGB-D against Nice-SLAM, Co-SLAM, ESLAM, BSLAM, and others, reporting average ATE RMSE of 0.45 cm on Replica (vs. 0.63 cm for ESLAM), Depth L1 of 0.89 cm, 66.86% completion within 1 cm, and PSNR of 31.62 dB, with a claimed 25% reduction in depth L1 error over the prior state of the art, together with a runtime of about 8.4 FPS on Replica. The supplement includes per-scene tables, ablations (representation choice, reweighting terms, BA components, hash grid sizes, Gaussian-assumption uncertainty), a telescoping-sum proof for the termination probability, and a plug-in experiment showing the strategic BA transfers to BSLAM.","tokens_in":28729,"tokens_out":20468,"duration_ms":169110,"significance":"If the reported results hold up, Uni-SLAM is a meaningful advance: it is the first dense implicit RGB-D SLAM system to use the rendering termination probability as a free, model-free confidence signal, and its benchmark numbers (best average ATE on all three datasets among the compared implicit methods) support the headline claims. Strengths worth crediting explicitly: baseline results are reproduced from open-source code with five runs (median reported) under a consistent mesh-culling protocol; the termination-probability identity is derived correctly as a telescoping sum; ablations substantiate the decoupled representation, the choice of reweighting terms, and the value of LLCO; and the BSLAM plug-in experiment demonstrates transfer of the BA strategy.","major_comments":[{"comment":"The central mechanism claim -- that beta_m = (1 - p(r))^2 identifies unreliable observations and that hard masking via Eq. (7) \"progressively filters out outliers\" -- is validated only in the regime where the model is uncertain because of missing coverage (invalid depth in Fig. 3, unseen areas in Supp. Fig. 10). The complementary regime is untested: after a pose perturbation, depth-guided samples can land on a previously learned surface, giving p(r) near 1 and CF_m = 1, so the inconsistent observation is up-weighted instead of filtered, which can amplify drift in the joint tracking/mapping loop. The supplement proof (Sec. A.2) establishes only the identity p(r) = 1 - exp(-sum sigma); it does not establish the semantic claim that p(r) is large exactly when the observation is reliable under pose error. I request a controlled experiment: inject pose perturbations of increasing magnitude during tracking (with the map frozen), report the resulting distribution of beta_m and the post-optimization tracking error, and compare against a baseline that masks only pixels with explicitly invalid depth. This would show whether the reweighting provides information beyond a standard depth-validity mask and whether the aggregate gains in Fig. 9 arise from the proposed mechanism.","section":"Sec. 3.2, Eqs. (4)-(7); Fig. 3; Sec. 4.3"},{"comment":"Several main-text numbers do not reconcile with the per-scene supplement tables. (i) The per-scene Depth L1 values for Ours in Tab. 14 (0.81, 0.77, 1.13, 0.70, 1.11, 1.52, 1.15, 0.99) average to 1.02 cm, yet both Tab. 14 and Tab. 2 report 0.89 cm; the headline claim of a 25% reduction in Depth L1 over ESLAM (1.18 cm) depends on this figure. (ii) The ESLAM completion ratio at 1 cm is 62.25% as the mean of the Tab. 14 per-scene values but is reported as 63.99% in Tab. 2. (iii) For Ours on office-3, Fig. 6 reports a 50.19% completion ratio at 1 cm while Tab. 14 lists 50.91%. Please reconcile these numbers or state clearly which runs or configurations each table refers to.","section":"Supp. Tab. 14; Tab. 2; Fig. 6"},{"comment":"The quantitative comparison is statistically asymmetric: for the baselines the paper reports the median of five runs, while the proposed method is reported as a single run without error bars. Some of the claimed margins are modest (TUM average ATE: 2.05 vs. 2.14 for ESLAM; completion ratio at 1 cm on Replica: 66.86 vs. 63.99), so the claim that Uni-SLAM \"significantly improves\" over current methods is not yet substantiated. Please report mean +/- std (or per-run values) for the proposed method over multiple runs for the main tables, and mark which differences are robust to run-to-run variation.","section":"Sec. 4.1, Tabs. 1-4"},{"comment":"The uncertainty pipeline relies on several hand-set, dataset-specific thresholds: beta_unc_m = 1e-2 for pixel masking, beta_unc = 1e-3 (2e-3 on TUM) for activating local BA, mapping frequency n = 4 (Replica, TUM) or 5 (ScanNet), and tau_cov = 0.95 for loop closure. Because Eq. (7) is a hard 0/1 mask, behavior is discontinuous in beta_unc_m, and no sensitivity analysis is provided. I ask for a sweep over these thresholds (at least beta_unc_m and beta_unc) on one Replica and one TUM sequence, and for a statement of which thresholds were used in the BSLAM plug-in experiment (Tab. 10), to establish that the reported gains do not depend critically on tuning.","section":"Supp. Sec. A.1; Sec. 3.4"}],"minor_comments":[{"comment":"In the MIPS-Fusion row, the Off-2 entry reads \"2020\", which appears to be a typo (likely 2.02 or 20.20); please correct it.","section":"Tab. 1"},{"comment":"The tables label the BSLAM baseline as \"BSLAM [47]\" in several places (e.g., Tab. 5 and Tab. 12); BSLAM is reference [20], while [47] is Point-SLAM.","section":"Tabs. 5 and 12"},{"comment":"In the Replica block, the Nice-SLAM mapping column reads \"29.3 x 0\", which appears to be a typo for \"29.3 x 60\" (cf. Tab. 5).","section":"Supp. Tab. 12"},{"comment":"The notation CF_m = 1(1 - beta_m) is confusing; write the indicator function explicitly, e.g., CF_m = 1{beta_m <= beta_unc_m}.","section":"Eq. (7)"},{"comment":"\"never unobserved regions\" should read \"never observed regions\".","section":"Sec. 3.2"},{"comment":"The figure legends contain a leftover template string (\"SA T [8] BUTD-DETR [8] ViewRefer [8]\") that should be removed.","section":"Supp. Figs. 17 and 19"},{"comment":"The comment \"Optimazation\" should be corrected to \"Optimization\".","section":"Algorithm 1"},{"comment":"The claims of \"real-time performance\" should be qualified: per Supp. Tab. 12 the method runs at 8.37 FPS on Replica, 4.88 FPS on ScanNet, and only 2.7 FPS on TUM RGB-D, which is below the real-time threshold.","section":"Abstract and Conclusion"},{"comment":"The abstract's statement that previous methods \"do not consider varying quality of input RGB-D data\" is too strong given that UNCLE-SLAM [48] is cited in the related work; please soften or qualify it.","section":"Abstract and Related Work"},{"comment":"The reweighting ablation reports results on \"TUM RGB-D\" without identifying the specific sequence; please specify which sequence(s) were used.","section":"Fig. 9"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent systems contribution and the benchmark results are likely to be real, but the requested revisions (failure-mode validation for the uncertainty signal, statistical reporting, threshold sensitivity, and reconciling main-text with supplement numbers) go beyond copy-editing. I would also ask the editor to have the authors verify Tab. 1 and Supp. Tab. 14 against their raw logs, since the MIPS-Fusion \"2020\" entry and the mismatch between the per-scene Depth L1 values and the reported 0.89 cm average suggest possible data-entry errors. The leftover template string in the Supp. Figs. 17/19 captions is a presentation red flag indicating the supplement was assembled from multiple drafts; a careful proofread is warranted. The novelty claim relative to UNCLE-SLAM is defensible but should be sharpened to emphasize what the model-free formulation adds over learned aleatoric uncertainty."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a well-engineered systems paper with a genuinely new uncertainty heuristic, and it deserves a real referee. The central result—consistently better tracking and reconstruction than ESLAM, Co-SLAM, BSLAM on Replica, ScanNet, and TUM—is supported by careful experiments. The main soft spot is that the uncertainty measure is never stress-tested under the failure mode where it could backfire.\n\nWhat's new: β=(1-p(r))^2, a model-free predictive uncertainty taken from the accumulated termination probability of volume rendering, used both to reweight the per-pixel losses and to trigger extra local BA via an image-level average. That particular combination, with decoupled hash grids for geometry and appearance, is not in the prior SLAM literature. The uncertainty needs no extra network or training, which is a real practical advantage. The paper earns credit for the BSLAM plug-in experiment—it shows the strategy transfers—and for the ablations that justify most design choices. The per-scene tables and consistent mesh culling make the comparison believable. The supplement's telescoping-sum proof is correct, though it formalizes the standard p(r)=1−exp(−Σσ) identity.\n\nSoft spots, in proportion. The stress-test note has a valid point. The paper assumes high β indicates unreliable data and low β indicates reliable data, but it validates that mainly with Fig. 3 and benchmark-level ablations. If the pose is already wrong, a ray sampled through the estimated depth can terminate in a previously learned surface with high accumulated probability; then β is low, CF=1, and the bad observation is up-weighted. That is exactly the situation where an outlier filter should help most, and the paper never analyzes it. The per-dataset thresholds (1e-2, 1e-3, 2e-3 on TUM) reinforce the concern that the heuristic is tuned rather than validated as a general data-quality measure. This is a real gap, but it does not sink the paper: the gains are consistent, the plug-in works, and the reweighting clearly helps with invalid depth. I would call it a missing robustness analysis rather than a fatal flaw.\n\nOther issues are minor. No error bars for the proposed method despite 5-run baselines. \"Real-time\" is sequence-dependent: 8.37 FPS on Replica is fine, 2.7 FPS on TUM is not really real-time. The first pose is initialized from ground truth (Algorithm 1), which is disclosed but never analyzed. Code is not released; for a system paper that matters.\n\nWho it's for: anyone working on dense neural RGB-D SLAM, especially on uncertainty-based loss weighting or BA scheduling. It is a solid within-subfield advance, not a paradigm shift. I'd send it to review and would cite it. Ask for code, error bars, and an explicit test of the uncertainty under artificial pose perturbation before accepting.","headline":"Solid, well-engineered SLAM systems paper with a genuinely new uncertainty heuristic; the main gap is that the heuristic is never stress-tested under pose failure.","tokens_in":29276,"tokens_out":2851,"would_cite":true,"duration_ms":27234,"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":"Uni-SLAM shows that weighting each pixel by a volume-rendering confidence score improves neural implicit SLAM tracking and mapping without giving up real-time operation.","keywords":["neural implicit SLAM","dense RGB-D reconstruction","predictive uncertainty","volume rendering termination probability","hash grid scene representation","bundle adjustment","indoor scene reconstruction","real-time mapping"],"falsifier":"Use a Replica or ScanNet sequence with known invalid-depth regions, run Uni-SLAM, and record per-pixel $p(r)$, $\\beta_m$, and depth error against the sensor depth. The claim predicts that pixels with $\\beta_m$ below threshold have small depth error and that high-$\\beta_m$ pixels coincide with invalid or unobserved depth; if any low-uncertainty pixel shows depth error larger than the truncation distance, or if the uncertainty map does not flag deliberately corrupted depth regions, the confidence proxy is falsified.","tokens_in":28281,"feed_emoji":"🗺️","tokens_out":8389,"duration_ms":73746,"temperature":0.7,"pith_summary":"Uni-SLAM sets out to show that a dense neural implicit SLAM system should not treat all RGB-D frames and pixels as equally trustworthy. Its central proposal is a cheap, model-free predictive uncertainty computed from the volume-rendering termination probability, which reweights the tracking and mapping losses so that poorly observed pixels count for less. The same image-level uncertainty triggers additional local bundle adjustment when the camera enters unfamiliar territory, instead of relying only on fixed-frequency global mapping. On the Replica, ScanNet, and TUM RGB-D benchmarks the paper reports state-of-the-art tracking accuracy and reconstruction quality among the compared neural implicit systems, with a 25% lower depth L1 error and a 66.86% sub-centimeter completion ratio on Replica, while running at about 8 FPS. If correct, this establishes uncertainty-aware selective optimization, rather than uniform fixed-frequency mapping, as the better design for real-time implicit indoor SLAM.","feed_headline":"Neural SLAM that weighs pixels by confidence cuts depth error 25%","feed_subtitle":"Uni-SLAM uses volume-rendering ray confidence to reweight losses and trigger local bundle adjustment, at 8+ FPS.","key_machinery":"Central machinery is the termination-probability confidence signal. Along a ray with sample weights $w_i = T_i(1-\\exp(-\\sigma(x_i)))$, the accumulated value $p(r)=\\sum_{i=1}^N w_i$ telescopes to $1-\\exp(-\\sum_i\\sigma(x_i))$, bounding it in $(0,1)$ by construction. The paper treats $p(r)$ as the probability that the ray terminates on a known surface, defines $\\beta_m=(1-p(r_m))^2$ for each pixel and $\\beta=\\frac{1}{M}\\sum_m\\beta_m$ per image, and converts $\\beta_m$ into a binary mask $CF_m=\\mathbf 1[\\beta_m\\le \\beta^\\text{unc}_m]$ that multiplies the SDF, depth, and tracking color losses. This is what lets the system discard outliers and emphasize reliable data during pose optimization. The second load-bearing mechanism is decoupled geometry and appearance hash grids, which allow high-frequency color and geometry to be sampled at different resolutions, plus the uncertainty-gated local-to-global bundle adjustment schedule that adds local BA when $\\beta>\\beta^\\text{unc}$ and LLCO when co-visibility exceeds 0.95.","core_discovery":"The paper's claim is that the accumulated termination probability of a sampled ray, $p(r)=\\sum_i w_i = 1-\\exp(-\\sum_i \\sigma(x_i))$, is a usable online measure of how well the current model knows the observed surface: near 1 for well-observed, accurately tracked pixels and near 0 for invalid-depth, occluded, or never-seen regions. From it Uni-SLAM derives pixel-level uncertainty $\\beta_m=(1-p(r_m))^2$ and image-level uncertainty $\\beta=\\frac{1}{M}\\sum_m \\beta_m$, uses a thresholded confidence mask $CF_m$ to reweight SDF, depth, and tracking color losses, and leaves the mapping color loss unweighted so RGB can compensate for invalid depth. A decoupled pair of multiresolution hash grids represents geometry and appearance separately, and image-level uncertainty plus a co-visibility check activates local bundle adjustment and loop-closure optimization on top of constant global BA. The reported outcome is improved pose estimation in every Replica scene, lower ATE on ScanNet and TUM RGB-D than the compared neural implicit SLAM systems, thinner structures in the reconstructed meshes, and real-time operation at roughly 8.4 FPS.","pith_inferences":["Editorial extension: because $\\beta$ is computed from the current model state, it could double as an online alarm for motion blur, sensor dropout, or aggressive camera motion, enabling the SLAM system to slow down, request new views, or refuse bad poses before drift accumulates.","Editorial extension: the same uncertainty could drive active next-best-view selection, since high-$\\beta$ rays mark regions the model has not yet observed; the paper mentions active sampling as future work but does not test it.","Editorial extension: the hand-set thresholds $\\beta^\\text{unc}_m$ and $\\beta^\\text{unc}$ vary across datasets, so a principled automatic calibration, such as normalizing by the per-image uncertainty distribution, would be needed for deployment outside the tuned benchmarks.","Editorial extension: the decision not to reweight color during mapping rests on RGB compensating for invalid depth; in a sensor with reliable depth, reweighting color in mapping might become beneficial, which could be tested directly."],"forward_implications":["If the results hold, a neural implicit SLAM system can gain accuracy by computing a per-pixel confidence from the model's own rendering weights, with no extra uncertainty network and no additional training cost.","Tracking on Replica improves to a 0.45 cm average ATE RMSE, with ScanNet and TUM RGB-D averages of 7.01 cm and 2.05 cm, both best among the neural implicit baselines compared.","Mapping quality on Replica improves to 0.89 cm depth L1, 0.92 cm accuracy and completion, and 66.86% completion within 1 cm, corresponding to thinner structures such as chair legs and table edges.","The uncertainty-driven mapping trigger is not tied to Uni-SLAM's own representation: applying it to BSLAM improves that system's ATE from 1.19 cm to 1.07 cm and its completion ratio from 57.18% to 58.36%.","Real-time operation is preserved: roughly 8.4 FPS on Replica with an RTX 4090, comparable to Co-SLAM and faster than ESLAM, Nice-SLAM, and BSLAM in the reported setup."],"supporting_citations":[{"why":"supplies the NeRF volume-rendering weights whose sum defines the termination probability p(r).","marker":"[36]"},{"why":"supplies the multiresolution hash encoding used for decoupled geometry and appearance grids.","marker":"[37]"},{"why":"motivates the (1-p)^2 predictive uncertainty that the paper adapts to a single real-time model.","marker":"[58]"},{"why":"provides the depth-guided volumetric rendering scheme and a principal baseline for tracking and reconstruction comparisons.","marker":"[77]"},{"why":"provides depth-guided ray sampling and pose optimization conventions and is a key baseline throughout the experiments.","marker":"[62]"},{"why":"is the main state-of-the-art baseline on Replica and ScanNet for both tracking accuracy and reconstruction metrics.","marker":"[22]"},{"why":"is a baseline system that also receives the proposed uncertainty-guided strategic BA in the plug-in experiment.","marker":"[20]"},{"why":"is the learnable Gaussian-assumption uncertainty baseline compared in ablations to justify the model-free definition.","marker":"[16]"}],"fun_headline_variants":["Uncertainty-aware rays sharpen neural SLAM's thin structures","Confidence-weighted losses cut depth error 25% in indoor SLAM","Real-time neural SLAM with ray confidence for fine details","Knowing what it sees helps neural SLAM map thin structures","Ray-based uncertainty improves neural SLAM accuracy and speed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system trusts that the accumulated termination probability is near 1 precisely for well-observed, accurately posed pixels and near 0 for invalid or unseen regions, so a confidently wrong model could be reweighted up instead of filtered.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty-aware rays sharpen neural SLAM's thin structures","Confidence-weighted losses cut depth error 25% in indoor SLAM","Real-time neural SLAM with ray confidence for fine details","Knowing what it sees helps neural SLAM map thin structures","Ray-based uncertainty improves neural SLAM accuracy and speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1498,"prompt_tokens":1015,"completion_tokens":483,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":399}},"tokens_in":631,"tokens_out":483,"duration_ms":5079,"temperature":1.0,"reasoning_tokens":399,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:34:20.116052+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use a Replica or ScanNet sequence with known invalid-depth regions, run Uni-SLAM, and record per-pixel $p(r)$, $\\beta_m$, and depth error against the sensor depth. The claim predicts that pixels with $\\beta_m$ below threshold have small depth error and that high-$\\beta_m$ pixels coincide with invalid or unobserved depth; if any low-uncertainty pixel shows depth error larger than the truncation distance, or if the uncertainty map does not flag deliberately corrupted depth regions, the confidence proxy is falsified.","supporting_citations":[{"cited_title":"Nerf: Representing scenes as neural radiance fields for view synthesis","cited_arxiv_id":null,"evidence_quote":"supplies the NeRF volume-rendering weights whose sum defines the termination probability p(r)."},{"cited_title":"Instant neural graphics primitives with a multiresolution hash encoding","cited_arxiv_id":null,"evidence_quote":"supplies the multiresolution hash encoding used for decoupled geometry and appearance grids."},{"cited_title":"Density-aware nerf ensembles: Quantifying predictive uncertainty in neural radiance fields","cited_arxiv_id":null,"evidence_quote":"motivates the (1-p)^2 predictive uncertainty that the paper adapts to a single real-time model."},{"cited_title":"Nice-slam: Neural implicit scalable encoding for slam","cited_arxiv_id":null,"evidence_quote":"provides the depth-guided volumetric rendering scheme and a principal baseline for tracking and reconstruction comparisons."},{"cited_title":"Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam","cited_arxiv_id":null,"evidence_quote":"provides depth-guided ray sampling and pose optimization conventions and is a key baseline throughout the experiments."},{"cited_title":"Eslam: Efficient dense slam system based on hybrid representation of signed distance fields","cited_arxiv_id":null,"evidence_quote":"is the main state-of-the-art baseline on Replica and ScanNet for both tracking accuracy and reconstruction metrics."},{"cited_title":"Benchmarking implicit neural representation and geometric rendering in real-time rgb-d slam","cited_arxiv_id":null,"evidence_quote":"is a baseline system that also receives the proposed uncertainty-guided strategic BA in the plug-in experiment."},{"cited_title":"Bayes’ rays: Uncertainty quantification for neural radiance fields","cited_arxiv_id":null,"evidence_quote":"is the learnable Gaussian-assumption uncertainty baseline compared in ablations to justify the model-free definition."}],"review_version":1}