{"id":"23121f22-ae37-44e0-8bf0-fb2a0bf106b7","arxiv_id":"2411.08279","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A dense RGB-D SLAM system that models camera motion during exposure and re-blurs rendered images, improving tracking and mapping on motion-blurred and sharp video.","lead":"MBA-SLAM is a camera tracking and 3D mapping system designed to keep working when video frames are blurry, by estimating how the camera moved during each exposure. It combines this motion-blur model with NeRF and Gaussian Splatting scene representations, and reports large accuracy gains on blurry benchmark sequences.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (7) does not follow from Eq. (5): the implemented blur model decouples rotation and translation, so the 'physical image formation process' claimed in the abstract is not the SE(3) interpolation actually used in tracker and mapper.","rationale":"The reader's weakest-assumption analysis targeted the external validity of the linear SE(3) trajectory model: real hand-held motions can change direction inside the exposure. My concern is internal and more specific: even under the paper's two-pose model, Eq. (7) is not the translation component of Eq. (5). This matters because Eqs. (24) and (26) use the implementable form, so the synthesized blur does not correspond to the physical model the paper claims. The error is first-order in the product of angular and translational displacement, which is large for the severe shakes in the synthetic blur benchmark. I do not see this as a rejection: the method is a substantial engineering contribution, the ablations support the tracker's value, and the empirical results are strong on blurry data. The correct remedy is to either implement the exact SE(3) exponential in Eq. (5) or explicitly present Eqs. (6)-(7) as an approximate decoupled model, with a quantitative estimate of the induced blur error. This reinforces, rather than overturns, the reader's conditional verdict.","tokens_in":29253,"tokens_out":8521,"duration_ms":97556,"concrete_test":"Directly test the claimed decomposition: set T_start = I, T_end = (RotZ(30°), t = [0, 0.1, 0]^T). Evaluate Eq. (5) at t = τ/2 using the exact SE(3) exponential, and compare the resulting translation with Eq. (7). Project a point at depth 1 m under both intermediate poses; if the pixel displacement exceeds 1 px, the implementation is not using Eq. (5). As a consequential check, replace Eq. (7) with the exact translation from Eq. (5) in the Gaussian Splatting mapper and rerun ArchViz-1; if ATE or PSNR changes materially, the mismatch is not negligible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2.2 defines the intra-exposure pose by Eq. (5): T_t = T_start · exp((t/τ) · log(T_start^{-1} T_end)). It then says Eqs. (6)-(7) decompose this into quaternion interpolation plus linear translation, with Eq. (7) reading t_t = t_start + (t/τ)(t_end - t_start). This is not a decomposition of Eq. (5). Writing G = T_start^{-1} T_end = (R_rel, t_rel), the translation of exp((t/τ) log G) is generally not (t/τ) t_rel; it is the translation of a constant body-frame twist, which contains rotation-translation coupling terms. For example, a 30° rotation combined with a lateral translation produces a non-zero translation component along the rotation axis's cross-product direction at mid-exposure, while Eq. (7) gives zero. The curve actually implemented in Eqs. (24) and (26) is therefore not a one-parameter subgroup of SE(3) and does not correspond to constant body-frame velocity. Since both the tracker and the mapper synthesize blur by averaging virtual views along this curve, the central claim of 'accurately modeling the physical image formation process' is only approximately true, with the largest error in precisely the combined rotation-plus-translation shakes used in the ArchViz experiments. The joint optimization can partially absorb the mismatch by adjusting T_start and T_end, but the fitted trajectory and deblurred map are then biased by the unmodeled inconsistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"MBA-SLAM proposes a dense RGB-D SLAM system for motion-blurred video. The key idea is to represent the camera motion during each frame's exposure by two poses, T_start and T_end, and to synthesize the observed blur by averaging virtual sharp views rendered at interpolated poses along the trajectory. The system has a CUDA-implemented blur-aware tracker that re-blurs a rendered keyframe and aligns it to the current blurry frame, and a mapper that jointly optimizes the scene representation (either tri-plane NeRF or 3DGS) and the intra-exposure trajectories. The paper evaluates on the synthetic ArchViz blur dataset, selected blurry ScanNet/TUM sequences, a newly captured RealSense+Vicon blur dataset, and sharp Replica/ScanNet/TUM benchmarks. The reported ATE, PSNR/SSIM/LPIPS, and mesh reconstruction results are strong on the blur datasets, and ablations show that the blur-aware mapper and the frame-to-frame tracker each improve performance. The method explicitly extends the authors' earlier MBA-VO, BAD-NeRF, and BAD-Gaussians work into a complete SLAM pipeline with a faster tri-plane NeRF backend.","tokens_in":29559,"tokens_out":10845,"duration_ms":105053,"significance":"If correct, the paper would show that dense RGB-D SLAM can operate directly on severely blurred video by jointly recovering the scene and the intra-exposure trajectory, avoiding the usual front-end deblurring or pose-failure cascade. The empirical support is substantial: on ArchViz the proposed GS version reaches average ATE 0.84 cm and PSNR 28.82 dB, far exceeding the compared baselines; the ablations isolate the contribution of the blur model and of the new tracker; and the authors commit to releasing code and introduce a real blur dataset with motion-capture ground truth. The central caveat is that the motion-model derivation contains an internal inconsistency (Eq. (7) is not the translation component of the SE(3) interpolation in Eq. (5)), so the 'physical image formation' claim needs to be repaired or softened. In addition, the abstract's claim to surpass all prior methods is not supported on the sharp Replica and TUM datasets.","major_comments":[{"comment":"The motion model is not what it is claimed to be. Eq. (7) states t_t = t_start + (t/τ)(t_end − t_start), and the text says it is a decomposition of Eq. (5), but the translation component of exp((t/τ) log(T_start^{-1} T_end)) is only equal to that linear expression for pure translation. For a general relative transform (R_rel, t_rel), the exponential-map translation contains rotation–translation coupling terms (e.g., terms proportional to [ω]v via the left Jacobian of SO(3)); these are absent from Eq. (7). Because the tracker's re-blurring in Eq. (24) and the Jacobians in Eqs. (10)–(12) and (22)–(23) are built from Eq. (7), the implemented trajectory is a decoupled rotation-plus-linear-translation interpolation, not constant-velocity SE(3) motion. The abstract's phrase 'accurately modeling the physical image formation process' is therefore an overstatement. Please either implement the true SE(3) interpolation and update the Jacobians, or explicitly present Eq. (7) as an approximation and reword the physical-model claims.","section":"Sec. 3.2.2, Eqs. (5)–(7)"},{"comment":"Eq. (26) restates the virtual pose as the full SE(3) exponential T_start exp((i/(n−1))τ log(T_start^{-1} T_end)), while the Jacobian derivations in Sec. 3.2.2, particularly Eqs. (10)–(12) and (22)–(23), are valid only for the decoupled model of Eq. (7). These two models are not equivalent. The manuscript should specify unambiguously whether the tracker and the mapper use Eq. (26) or Eqs. (6)–(7); if the implementation uses the decoupled model, Eq. (26) and any downstream uses must be changed, and the effect of the mismatch on blur synthesis should be discussed.","section":"Sec. 3.2.4, Eqs. (24) and (26)"},{"comment":"The abstract and introduction claim that MBA-SLAM 'surpasses previous state-of-the-art methods in both camera localization and map reconstruction' without qualification. This is contradicted by the paper's own sharp-dataset results: in Table 4, RTG-SLAM has average ATE 0.18 cm on Replica, substantially lower than Ours-GS 0.35 cm and Ours-NeRF 0.41 cm; in Table 7, on TUM RGB-D, RTG-SLAM (1.06 cm), Photo-SLAM (1.28 cm), and MonoGS (1.47 cm) all outperform Ours-GS (1.98 cm). The text in Sec. 4.3 acknowledges 'except RTG-SLAM', but the abstract and contribution list do not. Please restrict the SOTA claim to the motion-blur datasets, or rephrase it as 'among NeRF-based and 3DGS-based methods' with the specific exceptions stated.","section":"Abstract, Sec. 4.3, Tables 4 and 7"}],"minor_comments":[{"comment":"The self-captured Realsense dataset is not described with exposure times, blur magnitudes, motion speeds, or synchronization details; please add these to support the dataset as a community resource.","section":"Sec. 4.1"},{"comment":"The row 'Photo-SLAM [16]' in Table 8 uses the reference number of MonoGS, and Table 9's 'PointSLAM [12]' uses ESLAM's number; please correct the table citations.","section":"Tables 8 and 9"},{"comment":"Many baseline cells are marked with ✗ or ✖; the captions should state the exact failure criterion (e.g., tracking divergence threshold or code error) and whether these runs were excluded from the averages.","section":"Tables 2 and 3"},{"comment":"The text says all methods were run five times and averaged, but no standard deviations are reported; given the large run-to-run spread visible in some baselines (e.g., Point-SLAM on ArchViz), adding variance or per-run values would improve the comparison.","section":"Tables 1 and 4"},{"comment":"The rationale for n=13 is saturation, but ATE on ArchViz-3 keeps improving from 1.413 cm at n=13 to 1.223 cm at n=17; please state the trade-off criterion more explicitly.","section":"Sec. 4.5, Table 12"},{"comment":"The depth loss D(x) uses the middle pose while the color loss B(x) integrates over the whole trajectory; one sentence explaining why depth is not blurred in the same way would avoid confusion.","section":"Sec. 3.3.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental integration of the authors' own prior components (MBA-VO, BAD-NeRF, BAD-Gaussians) with a tri-plane speed-up. The disclosure is explicit and the experiments are extensive, which makes the contribution acceptable as a journal paper if the motion-model inconsistency and the overclaims are fixed. I would not recommend rejection based on novelty alone; the empirical study on blurry RGB-D SLAM is valuable. One point to monitor is the absence of standard deviations in the five-run averages, especially for blur datasets where failures are common."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Peng and coauthors have built a genuinely useful system: MBA-SLAM couples a blur-aware tracker with NeRF and 3DGS mappers, estimates per-exposure start/end poses, and re-blurs rendered images to match captured blur. No COLMAP initialization, no offline deblurring, and the tri-plane NeRF backend gives a real speedup over BAD-NeRF. The ablations are honest: turning on the blur model and swapping in the tracker helps on both synthetic and real blurry data, and on sharp Replica sequences the system is competitive even without blur modeling. That deserves credit.\n\nThe paper overstates itself in two ways. First, the abstract and method claim that the blur model comes from the physical image formation process, and Eq. (5) defines the trajectory as an SE(3) geodesic. But Eqs. (6)-(7) decompose this into quaternion slerp plus linear translation, which is not the same curve unless the rotation is zero. The exponential map of a twist couples rotation and translation; the translation of exp((t/τ)logG) is not linear interpolation. So the implemented model—used by both tracker and mapper—is a rotation-translation-decoupled approximation. For short exposures the difference is small, and the joint optimization can compensate, which is probably why the experiments still work. But the claim of 'accurately modeling the physical image formation process' should be walked back or the implementation should actually use the SE(3) exponential.\n\nSecond, the sharp-dataset claims are too strong. On Replica, RTG-SLAM has lower ATE, and MonoGS/Photo-SLAM beat their PSNR; the paper's own tables show this, yet the text says 'superior or comparable.' That's not the end of the world, but the framing needs to align with the numbers.\n\nMinor but real issues: they report averaging five runs but offer no error bars, the three public blur sequences are not identified, and the self-captured dataset promised in the conclusion is not released. The code is public, which is good.\n\nIf the authors fix Eq. (7) (or relabel it as an approximation), release the dataset, and add variance bars, this becomes a solid contribution. It deserves a serious referee.","headline":"Solid system paper on blur-robust dense SLAM, but the flagship 'physical formation' claim is undercut by an inconsistency between the SE(3) interpolation and its translation/rotation decomposition.","tokens_in":30164,"tokens_out":2474,"would_cite":true,"duration_ms":25732,"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":"MBA-SLAM claims that dense RGB-D SLAM can track and map from severely blurred video by modeling each blurred frame as an average of sharp views along the exposure trajectory, jointly optimizing trajectory and scene.","keywords":["motion blur","dense RGB-D SLAM","3D Gaussian Splatting","Neural Radiance Fields","bundle adjustment","camera trajectory estimation","image deblurring","visual odometry"],"falsifier":"Take a camera whose ground-truth motion inside each exposure contains an inflection point, for example a jerk-and-reverse trajectory with the same total displacement as a straight segment, and run the method on the resulting RGB-D sequence. If the per-frame trajectory estimates and deblurred renderings degrade substantially relative to a straight-line exposure of equal displacement, that is direct evidence that the linear-in-SE(3) model is the limiting factor; the expected signature is residual streak directions in the deblurred image that the average of linearly interpolated views cannot produce.","tokens_in":29013,"feed_emoji":"🎥","tokens_out":11245,"duration_ms":96456,"temperature":0.7,"pith_summary":"This paper sets out to establish that dense visual SLAM does not require sharp input frames: by inserting the physical motion blur formation process into both tracking and mapping, a stream of blurry RGB-D images can yield accurate camera trajectories and a sharp, high-fidelity 3D map. The method represents the camera motion during each exposure by two poses, one at the start and one at the end, with intermediate poses interpolated in $\\mathrm{SE}(3)$; blurred images are synthesized by averaging the sharp views rendered along that path and are compared against the captured frames. Because the synthesis is differentiable, the trajectory and the scene representation (a tri-plane radiance field or a set of 3D Gaussians) are optimized jointly, which turns blur into a source of constraint rather than a failure mode. On synthetic and real blurry datasets the reported localization and reconstruction results surpass earlier NeRF- and Gaussian-based SLAM systems, and the same pipeline remains competitive on standard sharp datasets when the blur model is switched off.","feed_headline":"Blur-aware SLAM out-tracks earlier dense SLAM methods","feed_subtitle":"Treats each blurred frame as an average of sharp views along the exposure, so localization and mapping work on blurry RGB-D video.","key_machinery":"The load-bearing mechanism is differentiable re-blurring. For each sampled image patch the tracker transfers pixels to virtual poses interpolated along the trajectory, reads intensities from the sharp reference by bilinear interpolation, averages them as in Eq. (24), and minimizes the difference from the captured blurry intensities. The trajectory itself is compactly parameterized by two $\\mathrm{SE}(3)$ poses, $T_{\\mathrm{start}}$ and $T_{\\mathrm{end}}$, connected by Lie-algebra interpolation (Eq. (5)); this small unknown set is what lets the tracker run at real-time speed and gives the mapper a well-posed bundle adjustment over the exposure intervals. The end-to-end differentiability of the averaging model lets gradients flow from the blur residual into both the scene representation and the per-frame trajectory, which is why the same formulation serves tracking, mapping, and deblurring.","core_discovery":"On its own terms, the central discovery is that a blurred frame can be treated as a physically differentiable average of sharp views: $B(\\mathbf{x}) \\approx \\frac{1}{n}\\sum_{i=0}^{n-1} I_i(\\mathbf{x})$, where each virtual view $I_i$ is rendered from a pose $T_t = T_{\\mathrm{start}} \\exp(\\frac{t}{\\tau}\\log(T_{\\mathrm{start}}^{-1}T_{\\mathrm{end}}))$ inside the exposure. Equipped with this model, the tracker aligns a rendered sharp keyframe to the current blurry frame by re-blurring, and the mapper jointly optimizes keyframe trajectories and scene parameters by minimizing photometric and depth residuals on the synthesized blur. The paper further claims that the resulting system outperforms previous state-of-the-art dense visual SLAM methods on motion-blurred data while also holding its own on sharp benchmarks, and that either a tri-plane NeRF or a 3D Gaussian Splatting backend can carry the mapping role.","pith_inferences":["Editorial inference: the linear trajectory model is the natural ceiling of the method; a sequence with a sharp direction reversal inside one exposure should expose residual blur that the two-pose model cannot explain, and replacing the interpolation with a spline or fusing inertial readings is the obvious extension.","Editorial inference: the re-blurring loss does not fundamentally require measured depth once a coarse scene exists, so a monocular version that renders depth from the map and then re-blurs could inherit some of the same tolerance to blur.","Editorial inference: if blurred frames become constraints instead of outliers, motion blur handling in SLAM may shift from preprocessing the frames to estimating a per-frame trajectory, with implications for rolling-shutter and hand-held capture where blur is frequent."],"forward_implications":["Dense SLAM systems could operate directly on video captured in low light or with long exposure, where current NeRF- and Gaussian-based pipelines typically lose tracking or produce corrupted maps.","The learned map is a sharp representation of the scene, so the same pipeline outputs deblurred renderings of the environment without a separate deblurring network.","Per-frame start and end poses provide temporal information about camera motion during each exposure, not just a single instantaneous pose.","Because the blur model is representation-agnostic, the approach can be ported to future differentiable scene representations by keeping the same re-blurring loss.","On sharp inputs, keeping the blur model off restores full speed while retaining competitive accuracy, so one system covers both sharp and blurry operation."],"supporting_citations":[{"why":"It supplies the motion-blur-aware visual odometry tracker and the synthetic ArchViz blur dataset used in the evaluation.","marker":"[18]"},{"why":"It establishes the bundle-adjusted deblurring formulation for radiance fields that the NeRF backend extends to tri-planes.","marker":"[22]"},{"why":"It establishes the bundle-adjusted deblurring formulation for Gaussian splatting that the 3DGS backend builds on.","marker":"[23]"},{"why":"It provides the tri-plane radiance field representation and the auxiliary free-space and SDF losses used by the NeRF mapper.","marker":"[12]"},{"why":"It provides the 3D Gaussian Splatting scene representation and differentiable rasterizer used by the GS mapper.","marker":"[8]"},{"why":"It provides the volumetric rendering paradigm on which the radiance field backend is based.","marker":"[7]"},{"why":"It is a primary Gaussian Splatting SLAM baseline whose tracking, rendering, and reconstruction results are compared against.","marker":"[15]"},{"why":"It provides the TUM RGB-D benchmark and the absolute trajectory error metric used for real-data evaluation.","marker":"[21]"}],"fun_headline_variants":["Blur-aware SLAM turns blurry video into sharp maps and poses","Treats blur as averaged sharp frames; SLAM learns to deblur","Motion blur no match: SLAM tracks and maps blurry scenes","New SLAM deblurs as it maps: blurs become detail"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that during one exposure the camera follows a straight path in six-degree-of-freedom motion between its start and end poses, so every intermediate view lies on a linear interpolation; if the camera changes direction inside the exposure, the synthesized blur no longer matches the captured blur and the joint optimization cannot fully repair the mismatch.","fun_headline_variants_meta":{"raw":{"variants":["Blur-aware SLAM turns blurry video into sharp maps and poses","Treats blur as averaged sharp frames; SLAM learns to deblur","Motion blur no match: SLAM tracks and maps blurry scenes","New SLAM deblurs as it maps: blurs become detail"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000818,"raw_usage":{"total_tokens":3626,"prompt_tokens":1035,"completion_tokens":2591,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":2512}},"tokens_in":651,"tokens_out":2591,"duration_ms":17267,"temperature":1.0,"reasoning_tokens":2512,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:46:09.384997+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a camera whose ground-truth motion inside each exposure contains an inflection point, for example a jerk-and-reverse trajectory with the same total displacement as a straight segment, and run the method on the resulting RGB-D sequence. If the per-frame trajectory estimates and deblurred renderings degrade substantially relative to a straight-line exposure of equal displacement, that is direct evidence that the linear-in-SE(3) model is the limiting factor; the expected signature is residual streak directions in the deblurred image that the average of linearly interpolated views cannot produce.","supporting_citations":[{"cited_title":"MBA-VO: Motion Blur Aware Visual Odometry,","cited_arxiv_id":null,"evidence_quote":"It supplies the motion-blur-aware visual odometry tracker and the synthetic ArchViz blur dataset used in the evaluation."},{"cited_title":"BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields,","cited_arxiv_id":null,"evidence_quote":"It establishes the bundle-adjusted deblurring formulation for radiance fields that the NeRF backend extends to tri-planes."},{"cited_title":"ESLAM: Efficient dense SLAM system based on hybrid representation of signed distance fields,","cited_arxiv_id":null,"evidence_quote":"It provides the tri-plane radiance field representation and the auxiliary free-space and SDF losses used by the NeRF mapper."},{"cited_title":"3D Gaus- sian Splatting for Real-Time Radiance Field Rendering,","cited_arxiv_id":null,"evidence_quote":"It provides the 3D Gaussian Splatting scene representation and differentiable rasterizer used by the GS mapper."},{"cited_title":"NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis,","cited_arxiv_id":null,"evidence_quote":"It provides the volumetric rendering paradigm on which the radiance field backend is based."},{"cited_title":"SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM,","cited_arxiv_id":null,"evidence_quote":"It is a primary Gaussian Splatting SLAM baseline whose tracking, rendering, and reconstruction results are compared against."},{"cited_title":"A benchmark for the evaluation of rgb-d slam systems,","cited_arxiv_id":null,"evidence_quote":"It provides the TUM RGB-D benchmark and the absolute trajectory error metric used for real-data evaluation."}],"review_version":1}