{"id":"30e5fa34-0263-4dcc-ac3f-9f0119d73df3","arxiv_id":"2501.13402","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"VIGS SLAM uses IMU preintegration as the initial guess for GICP tracking inside a 3D Gaussian Splatting SLAM pipeline, enabling lower drift and better rendering in large indoor scenes.","lead":"VIGS SLAM combines RGB-D and IMU data to make 3D Gaussian Splatting SLAM work in large indoor spaces, using IMU motion estimates to give the point-cloud tracker a better starting point. It reports much lower trajectory error than prior 3DGS SLAM systems on two large-scale indoor datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Suspected inconsistency in IMU preintegration equations (Eqs. 10-12) raises doubt about the claimed mechanism; a re-derivation or ablation is needed.","rationale":"The reader's verdict is CONDITIONAL, focusing on extrinsic and bias assumptions. My review agrees that the mechanism is plausible, but I found a more fundamental internal problem: the preintegration equations as written are inconsistent with the measurement model and with standard IMU preintegration. This is not merely a missing baseline or ablation; it affects the paper's central derivation. If Eqs. (10)-(12) are literal, the proposed method cannot literally be what is implemented, or the equations are wrong. Either way, the paper's central claim is not adequately supported. A re-derivation or a clarifying statement about implementation is required before acceptance. The concrete test above would settle which case applies. I therefore recommend REJECT, though with a path to resubmission if the derivation is corrected and verified.","tokens_in":9750,"tokens_out":1505,"duration_ms":12101,"concrete_test":"Independently re-derive the preintegration equations from the measurement model in Eqs. (5)-(6) using the standard manifold preintegration of Forster et al. (reference [18]). Specifically, write p_k_t+1 and v_k_t+1 in terms of a_t_hat and the gravity and bias terms, then derive the relative motion Delta-R, Delta-v, Delta-p. If Eq. (10)-(12) are missing the gravity term or bias handling, confirm by checking whether the predicted relative pose I_T_k-1_k changes direction when the gravity vector is flipped. Additionally, run the provided code (if released) on one uHumansV1 sequence with the gravity term artificially set to zero in Eqs. (10)-(11); if the ATE remains at roughly 35 cm, the equations are not the operative mechanism; if ATE degrades sharply, the paper should state the correct formulation.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim is that IMU preintegration provides a good initial guess for GICP, fixing tracking in large-scale scenes. However, the preintegration equations in Section III-B appear inconsistent with the measurement model in Eqs. (5)-(6) and with the state update in Eq. (16). Specifically, Eq. (5) defines the IMU measurement as a_t_hat = a_t + R^T g + b_a + n_a, but the preintegrated position/velocity terms in Eqs. (10)-(11) are written using only (a_t_hat - b_a), omitting the gravity term R^T g that Eq. (5) requires to be subtracted. If these equations are taken literally, the preintegrated relative pose would contain a gravity-induced error that grows quadratically with the time interval between frames. Moreover, the update in Eq. (16) resets position, rotation, and velocity, but not the biases b_a and b_g; the preintegrated measurements in Eqs. (10)-(12) also do not appear to use bias estimates from any estimator (e.g., a filter or factor graph). If the bias values are frozen or zero, the initial guess degrades over longer intervals. The reported large-scale results could still be correct if the implementation uses a different formulation or if the IMU intervals are short enough that gravity and bias errors remain small, but the paper does not state this. As written, the derivation is internally incomplete and the claimed mechanism is not verified by the equations.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VIGS SLAM, a 3D Gaussian Splatting SLAM system that fuses RGB-D and IMU measurements for large-scale indoor environments. The key idea is to use IMU preintegration between consecutive frames as the initial guess for GICP-based point-cloud tracking, which is claimed to prevent the tracking failures seen in GS-ICP SLAM in large, dynamic scenes. The mapping stage is a 3DGS back-end with photometric, SSIM, and depth losses. The experiments on uHumansV1 and uHumansV2 report ATE reductions from hundreds of centimeters to tens of centimeters compared with GS-ICP SLAM, with improved rendering metrics, and the authors claim parity with classical visual-inertial odometry and superiority over existing 3DGS SLAM methods.","tokens_in":10066,"tokens_out":6131,"duration_ms":754439,"significance":"If the claims hold, the paper offers a simple and potentially practical insight: a good IMU-preintegrated initial guess can stabilize GICP-based tracking for 3DGS SLAM in large dynamic environments, reducing ATE by an order of magnitude relative to GS-ICP SLAM on uHumansV1 while also improving rendering metrics. The use of external datasets and comparison with standard classical and 3DGS baselines is a strength, and the proposed mechanism is falsifiable in principle. However, the significance is currently tempered by the absence of an ablation separating the IMU contribution from other changes, the lack of comparison to the most closely related visual-inertial 3DGS baseline (MM3DGS), and equation-level inconsistencies in the preintegration derivation that make the mechanism not yet fully verified.","major_comments":[{"comment":"The preintegration definition is inconsistent with the kinematic model. Eq. (5) states that the accelerometer measurement is a_hat = a + wR wg + b_a + n_a, so Eq. (7) correctly subtracts the gravity term wR wg. However, the preintegrated terms alpha and beta in Eqs. (10)-(11) use only (a_hat - b_a) and drop the gravity term. If these preintegrated values are then used in Eqs. (13)-(14) to form the relative pose initial guess, a gravity-induced error that grows quadratically with the interval is introduced. Either the gravity term must appear in the relative pose formula (as in the standard preintegration model, e.g., Delta p = 1/2 g Delta t^2 + R_i alpha), or the notation must be clarified to show that the gravity contribution is handled separately.","section":"Section III-B, Eqs. (5)-(12)"},{"comment":"The update step resets position, rotation, and velocity but never estimates or resets the accelerometer and gyroscope biases b_a and b_g. Since the preintegrated measurements in Eqs. (10)-(12) depend on these biases, a frozen or zero bias assumption is required; the paper neither states this assumption nor demonstrates its validity on long intervals. This matters because the central claim is that the IMU preintegration provides an accurate initial guess for GICP. Please either add bias estimation or provide evidence that the intervals are short enough for bias drift to be negligible, and report sensitivity to the extrinsic parameter C_I T.","section":"Section III-C, Eq. (16)"},{"comment":"The evaluation does not yet support the load-bearing claim that the IMU preintegration, rather than other implementation changes, causes the improvement. There is no comparison to MM3DGS [9], which is the closest visual-inertial 3DGS method; there is no ablation that runs GS-ICP SLAM with VIGS's mapping and keyframe settings but without the IMU initial guess; and the reported ATE and PSNR values are single runs with no error bars or repeated-trial statistics. An ablation separating the IMU contribution is necessary to validate the mechanism described in Section III-A.","section":"Section IV and Tables I-III"},{"comment":"The optimization objective is written with the wrong sign. For Gaussian distributions, log p(d_m) is proportional to -1/2 d_m^T (Sigma_tgt + T Sigma_src T^T)^-1 d_m plus a constant, so maximizing the sum of d_m^T Sigma^-1 d_m is not equivalent to maximum likelihood; the correct formulation is an argmin over the Mahalanobis distances. This technical error in the derivation should be corrected.","section":"Section III-A, Eq. (4)"}],"minor_comments":[{"comment":"The claim that the method is 'the first to propose that Gaussian Splatting-based SLAM can be effectively performed in large-scale environments by integrating IMU sensor measurements' conflicts with the existence of MM3DGS [9], which is cited but not discussed in the introduction; please temper the novelty claim.","section":"Abstract and Introduction"},{"comment":"Equations (5) and (6) are labeled as 'corrected' measurements but include bias and noise terms; rename them as raw measurement models or clarify the wording.","section":"Section III-B, Eqs. (5)-(6)"},{"comment":"The rotation notation is ambiguous: wR, R_k^t, and the rotation in Eq. (13) are not all defined consistently. Please define the world frame, the IMU body frame, and the camera frame explicitly, and state which rotation transforms between which frames.","section":"Section III-B, notation"},{"comment":"The framework diagram includes 'Is Keyframe?' and 'Is Mapping-only Keyframe?' branches, but the criteria for keyframe selection are never given in the text; please add the keyframe selection rule.","section":"Fig. 2 and Section III-D"},{"comment":"The Gaussian definition in Eq. (17) omits the normalization constant; while this may be intentional for rendering, the mathematical expression as written is not a proper probability density, so please add a note or the constant.","section":"Section III-D, Eq. (17)"},{"comment":"There are many typographical and language errors ('breifly', 'intial', 'vy provinding', 'tracnsformation', 'higer', 'meausrements', 'significanty', and others); a careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper's main risk is novelty disclosure relative to MM3DGS, which is cited but not compared. If the authors can add the comparison and the ablation, and fix the preintegration equations, the contribution is publishable. The writing quality is below the standard for the venue but appears fixable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a straightforward, mostly sound engineering paper. The core idea is simple—use IMU preintegration to seed GICP registration instead of a constant-velocity or identity guess—and the reported gains are large and consistent with that mechanism. On uHumansV1, ATE drops from about 6.8 m for GS-ICP SLAM to 0.35 m for VIGS SLAM, and rendering metrics improve across the board. That is a real effect, and the paper should be taken seriously.\n\nWhat is actually new: the specific combination of IMU preintegration with GICP tracking in a 3DGS SLAM front-end. Prior work in the cited literature—MM3DGS—fused IMU with photometric tracking, and GS-ICP used GICP without IMU. So the technical contribution is a clean, sensible integration that extends 3DGS SLAM beyond room scale. The paper is also honest about being loosely coupled and lacking loop closure; the limitations section is not hiding anything.\n\nThe soft spots are real but addressable. The novelty claim is overstated: it says \"first\" for 3DGS SLAM with IMU, but MM3DGS already did IMU+3DGS, so that claim should be rephrased. More importantly, MM3DGS is not included as a baseline, which the reader will notice. There are no error bars or repeated runs, and key hyperparameters (keyframe selection, GICP voxel size, loss weights) are unspecified. The IMU preintegration equations in Section III-B are written imprecisely. The measurement model in Eq. (5) includes a gravity term, and the recursive equations (7–8) subtract it, but the preintegrated alpha and beta in Eqs. (10–11) do not show that subtraction. Taken literally, a long preintegration interval would accumulate a gravity error that grows quadratically. However, the method resets the preintegration at every frame and uses it only between consecutive frames, so the error is on the order of a few cm at typical frame rates—not a collapse. The paper does not say this, and the derivation should be fixed or clarified. Similarly, bias terms are not estimated or updated, which is fine for short intervals but should be stated and ideally ablated.\n\nWho this is for: people working on 3DGS SLAM or on fusing IMU with dense mapping front-ends. It is an engineering contribution, not a theoretical one, but it is useful and likely correct.\n\nRecommendation: send it to peer review. The empirical result is striking, the mechanism is plausible, and the gaps—proper derivation, MM3DGS comparison, error bars—are exactly what a good referee can require. It is not a desk reject.","headline":"A plausible engineering result—IMU preintegration as an ICP initial guess—makes 3DGS SLAM work in large scenes, but the novelty claim is overstated and the IMU derivation needs cleanup.","tokens_in":741,"tokens_out":931,"would_cite":false,"duration_ms":63487,"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":"VIGS SLAM claims that seeding GICP tracking with IMU preintegration lets 3D Gaussian Splatting SLAM scale to large indoor environments.","keywords":["3D Gaussian Splatting SLAM","IMU preintegration","Generalized ICP","RGB-D SLAM","visual-inertial odometry","large-scale indoor SLAM","radiance field mapping","sensor fusion"],"falsifier":"Run VIGS on an uHumansV1 sequence with the IMU signals replaced by a constant-velocity prior while keeping the Eq. (16) update; if the ATE remains near 35.35 cm on Humans12 the inertial prior is not the cause, whereas degradation toward the 677.87 cm GS-ICP baseline would confirm the paper's mechanism.","tokens_in":9543,"feed_emoji":"🧭","tokens_out":8812,"duration_ms":72852,"temperature":0.7,"pith_summary":"The paper tries to establish that the failure of 3D Gaussian Splatting (3DGS) SLAM in large-scale scenes is a front-end tracking problem, not a mapping limitation. Its proposed system, VIGS SLAM, fuses RGB-D with an IMU and uses IMU preintegration between consecutive frames as the initial guess for Generalized ICP tracking, so the tracker starts near the true motion instead of locking onto wrong correspondences. On the uHumansV1 office dataset (65 m × 65 m, with moving people), the method reduces ATE from 677.87 cm (GS-ICP SLAM) to 35.35 cm on Humans12, and on the office scene of uHumansV2 from 674.38 cm to 144.72 cm, while PSNR improves in every scenario. The paper argues this is the first demonstration that 3DGS SLAM can work in large-scale indoor environments by integrating IMU measurements into the ICP-based front end. A sympathetic reader would take the central claim to be: a good inertial initial guess is the difference between a photorealistic map and a skewed one.","feed_headline":"IMU seeding cuts 3DGS SLAM trajectory error from meters to centimeters","feed_subtitle":"Preintegrating IMU data into the GICP initial guess lets 3D Gaussian Splatting SLAM track reliably in 65-meter indoor scenes.","key_machinery":"The load-bearing object is the IMU preintegration triplet $(\\alpha^k_{k+1}, \\beta^k_{k+1}, \\gamma^k_{k+1})$ of Eqs. (10)–(12), which summarizes all IMU measurements between two RGB-D frames into a relative motion. That relative motion is transformed into camera coordinates and composed with the previous pose to give GICP's starting transformation (Eqs. (13)–(14)). Under the GICP framework, each point is modeled as a Gaussian with covariance from k-nearest neighbors, and the matching cost weighs correspondences by those covariances (Eqs. (1)–(4)); starting close to the optimum is what prevents the nearest-neighbor search from snapping to a wrong surface. After tracking, the optimized pose is fed back to reset the preintegration state (Eq. (16)), keeping the inertial prior from drifting over the run.","core_discovery":"VIGS SLAM claims that replacing the default initialization of GICP tracking with IMU preintegration is sufficient to make 3DGS SLAM accurate and scalable in large indoor environments. The relative transformation from IMU preintegration, converted to camera coordinates via the extrinsic $T^C_I$, seeds the correspondence search at each frame; only after GICP converges are the IMU position, velocity, and rotation reset to the optimized values (Eq. (16)). With this mechanism, VIGS is reported to outperform the GS-ICP SLAM baseline by a large margin—ATE 35.35 cm vs 677.87 cm on uHumansV1 Humans12, 25.03 cm vs 776.79 cm on Humans24, and 46.86 cm vs 973.49 cm on Humans60—and to approach VINS-Mono's trajectory accuracy while additionally producing a photorealistic 3DGS map with better PSNR, SSIM, and LPIPS than prior 3DGS SLAM systems. On the uHumansV2 office scene, ATE drops from 674.38 cm to 144.72 cm. The paper's explanation is that a good initial guess keeps GICP from converging to wrong point correspondences in low-overlap, dynamic scenes, and those correct poses are what let the Gaussian map render cleanly.","pith_inferences":["Editorial: the principle generalizes to any geometric tracker sensitive to its initialization—NDT, point-to-plane ICP, or feature-based odometry—so the core insight is not specific to Gaussian Splatting.","Editorial: because Eq. (16) resets only pose and velocity, not the accelerometer and gyroscope biases, slow bias drift is the predicted long-term failure mode; bias re-estimation or a tightly-coupled formulation would be the natural next step to suppress low-speed drift.","Editorial: a clean ablation would run VIGS with a constant-velocity initial guess instead of IMU preintegration on the same keyframe schedule; the ATE gap between that variant and VIGS would isolate how much of the improvement comes from the inertial prior itself.","Editorial: perturbing the extrinsic calibration $T^C_I$ on the same datasets should produce a graceful ATE increase back toward the GS-ICP baseline, providing a direct sensitivity test of the paper's mechanism."],"forward_implications":["3DGS SLAM can shed its room-scale requirement: because GICP with an inertial prior tolerates much larger spacing between keyframes than photometric-loss tracking, the number of stored keyframes and memory use drop substantially.","Tracking accuracy becomes the gating factor for photorealistic mapping: once poses are accurate, the 3DGS renderer produces clean images, including correctly excluding moving humans from the reconstructed map.","VIGS reaches trajectory accuracy comparable to traditional visual-inertial odometry (e.g., 35.35 cm vs VINS-Mono's 18.85 cm ATE on uHumansV1 Humans12) while also maintaining a photorealistic radiance-field map, shifting the usual accuracy-versus-realism tradeoff.","The front-end gain is orthogonal to back-end improvements, so adding loop closure and a tightly-coupled visual-inertial optimization—listed as future work—should stack with this initial-guess mechanism rather than replace it."],"supporting_citations":[{"why":"Defines 3D Gaussian Splatting, the map representation and rendering loss used by the mapping stage.","marker":"[2]"},{"why":"The GS-ICP SLAM baseline this work extends; supplies the ICP-tracking plus 3DGS mapping pipeline and the scale normalization.","marker":"[10]"},{"why":"VINS-Mono, the visual-inertial odometry baseline whose ATE the paper compares against to show parity.","marker":"[16]"},{"why":"IMU preintegration on manifold, the formulation behind Eqs. (10)–(12) that produces the initial guess.","marker":"[18]"},{"why":"Provides the uHumansV2 apartment and office scenes used to evaluate RGB-based performance.","marker":"[19]"},{"why":"Voxelized GICP, the point-cloud registration algorithm used for tracking with covariance-weighted distances.","marker":"[21]"},{"why":"uHumansV1, the 65 m × 65 m visual-inertial dataset used for the main large-scale evaluation.","marker":"[22]"}],"fun_headline_variants":["IMU preintegration makes 3DGS SLAM scale to large indoor scenes","3DGS SLAM gets large-scale boost from IMU preintegration","VIGS SLAM: IMU priors fix 3DGS tracking at scale","IMU initial guess turns 3DGS SLAM into large-scale contender","Preintegrated IMU seeds GICP for scalable 3DGS SLAM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gain rests on the IMU preintegration between consecutive frames being an accurate enough initial guess for GICP to find correct correspondences in large, dynamic scenes; this in turn assumes the camera–IMU extrinsic is accurate and the un-reset accelerometer and gyroscope biases stay small over each preintegration interval.","fun_headline_variants_meta":{"raw":{"variants":["IMU preintegration makes 3DGS SLAM scale to large indoor scenes","3DGS SLAM gets large-scale boost from IMU preintegration","VIGS SLAM: IMU priors fix 3DGS tracking at scale","IMU initial guess turns 3DGS SLAM into large-scale contender","Preintegrated IMU seeds GICP for scalable 3DGS SLAM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000328,"raw_usage":{"total_tokens":1879,"prompt_tokens":1036,"completion_tokens":843,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":738}},"tokens_in":652,"tokens_out":843,"duration_ms":6632,"temperature":1.0,"reasoning_tokens":738,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:58:47.361066+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run VIGS on an uHumansV1 sequence with the IMU signals replaced by a constant-velocity prior while keeping the Eq. (16) update; if the ATE remains near 35.35 cm on Humans12 the inertial prior is not the cause, whereas degradation toward the 677.87 cm GS-ICP baseline would confirm the paper's mechanism.","supporting_citations":[{"cited_title":"V oxelized gicp for fast and accurate 3d point cloud registration,","cited_arxiv_id":null,"evidence_quote":"Voxelized GICP, the point-cloud registration algorithm used for tracking with covariance-weighted distances."}],"review_version":1}