{"id":"1b4640e4-3bdb-4ad0-b55a-b287ffbc5a57","arxiv_id":"2411.19289","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Combining a modified SORT tracker with the Segment Anything Model cuts visual-inertial odometry trajectory error in dynamic, partially occluded scenes.","lead":"ADUGS-VINS adds a tracking and segmentation layer to visual-inertial odometry, detecting moving objects such as cars and pedestrians and removing their image feature points before pose estimation. The authors report lower trajectory error than existing VIO systems on public datasets and in a real outdoor test, which matters for robots and vehicles navigating crowded scenes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The masking stage removes every detected object regardless of motion state, so static features are discarded; no mask-quality or static-discard metric supports the claimed generalization.","rationale":"The reader's weakest assumption identified the same core vulnerability: the pose-estimation gains depend on the perception pipeline removing exactly the dynamic features and no static ones. I sharpen this by noting that the pipeline is not merely imperfect; it is structurally motion-agnostic. YOLOv11 class labels cannot distinguish a parked car from a moving car, so the method systematically discards static features whenever an object of a dynamic class appears. This makes the central claim of generalizable improvement contingent on scene composition: in environments with many stationary vehicles or people, the method could be worse than a VIO that keeps those features. The paper does not report the necessary diagnostics (mask IoU, static-feature removal rate, dynamic-feature retention rate), and the ablation design does not isolate the masking decision from the tracking enhancement. My concrete test uses VIODE's ground-truth motion to quantify this directly and to benchmark against an oracle mask, which would settle whether the remaining error is due to perception or to the VIO optimization itself. I agree with the reader's CONDITIONAL verdict; the concern is addressable through additional experiments and does not require rejection of the method's plausibility. The paper is clearly written and the pipeline is re-implementable, but the unpublished code, undisclosed lambda/beta values, and lack of repeated trials remain as stated by the reader, reinforcing the conditional status.","tokens_in":11393,"tokens_out":3484,"duration_ms":34478,"concrete_test":"Using VIODE's synthetic ground truth (which includes object motion and semantic labels), compute per-frame the false-positive rate (static feature points removed) and false-negative rate (dynamic feature points retained) of the ADUGS-VINS feature-rejection stage. Then rerun pose estimation with a motion-oracle mask that removes only truly moving objects; if the oracle improves ATE by more than 10% over ADUGS-VINS on the 'high' sequences, the perception pipeline is the limiting factor. Additionally, report the same static-discard metric on the OpenLORIS market sequence to assess generalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core mechanism in Section III-B removes any tracked feature point that falls inside a mask produced by YOLOv11 + Mobile SAM (Section III-A). Detection is class-based, so every detected vehicle, pedestrian, shopping cart, or similar object is masked and its features are discarded irrespective of whether that object is actually moving. In scenes where objects of these classes are stationary (parked cars, standing pedestrians, carts at rest), valid static feature points are permanently removed from bundle adjustment. This is not an edge case: VIODE contains parked vehicles, and the OpenLORIS market sequence contains shopping carts that may be static. The paper gives no quantitative measure of mask precision/recall, no count of static features incorrectly removed, and no check of how many dynamic features remain in the optimizer. The ablation in Figure 6 compares only with and without the enhanced SORT, not against a motion-informed oracle mask. Without such evidence, the headline gains in Table I may reflect a favorable false-positive/false-negative balance on the selected sequences rather than a generalizable dynamic-object rejection mechanism. Because the paper explicitly claims 'generalization and adaptability' across environments, the motion-agnostic masking policy is the load-bearing risk: it can degrade accuracy in scenes with many static objects of dynamic classes by throwing away constraints the VIO optimizer needs.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADUGS-VINS, a visual-inertial odometry system that combines YOLOv11 detection, an enhanced SORT tracker with an adaptive Kalman filter, and Mobile SAM promptable segmentation to identify and mask dynamic objects. Feature points falling inside the resulting masks are discarded before KLT tracking and bundle adjustment, with a compensation strategy to maintain a sufficient number of static features. The method is evaluated on the VIODE and OpenLORIS-Scene datasets and on a qualitative real-world sequence, reporting improved ATE RMSE over several published VIO baselines.","tokens_in":11751,"tokens_out":4644,"duration_ms":42744,"significance":"If the reported gains hold, the paper demonstrates a practical integration of a promptable foundation model into VIO for dynamic-object rejection, which is a relevant direction given the limited generalization of earlier semantic-segmentation-based approaches. The release of a real-world visual-inertial dataset is a useful contribution. However, the empirical support is currently incomplete: the key masking policy is not accompanied by any quantitative assessment of segmentation quality or false-positive rate, the OpenLORIS comparison partly relies on results from another paper, and the real-world demonstration lacks ground truth. These gaps make the central generalization claim harder to verify. The paper's core mechanism is plausible and the VIODE results are internally consistent, so the work is likely to be of interest to the VIO/dynamic-SLAM community once the evidence is strengthened.","major_comments":[{"comment":"The masking stage removes feature points for every detection of a dynamic class (vehicle, pedestrian, shopping cart, etc.) regardless of whether the object is actually moving. In scenes containing parked cars or stationary pedestrians, this policy discards useful static features and can degrade pose estimation. The paper provides no quantitative measure of mask precision/recall, no count of static features incorrectly removed, and no analysis of how many dynamic features remain in the optimizer. The ablation in Fig. 6 only compares the full method with and without the enhanced SORT, not against a motion-aware policy. Since the paper explicitly claims generalization and adaptability across environments, this motion-agnostic masking assumption is load-bearing and must be justified or mitigated.","section":"III-A and III-B"},{"comment":"The OpenLORIS comparison for several baseline methods is taken directly from the prior paper [37], as acknowledged in the caption. Because the experimental protocol, parameter settings, and underlying implementations are not identical, the reported improvements may reflect differences in evaluation conditions rather than algorithmic superiority. The claim of outperforming state-of-the-art methods on OpenLORIS is therefore not fully supported. Please re-run the baselines under the same protocol as ADUGS-VINS, or restrict the 'outperforms' claim to the VIODE experiments where the comparison is self-contained.","section":"IV-C, Fig. 7"},{"comment":"The adaptive measurement noise covariance model R = diag(beta * erf(lambda * delta_RMSE)) is introduced without a derivation or a discussion of its properties. In particular, erf(0)=0 is a degeneracy: if the residual RMSE is zero, the filter assumes zero measurement noise, which is unrealistic and can lead to overconfidence. The values of lambda and beta are not reported, and the ablation in Fig. 6 does not isolate the effect of this adaptive rule versus a standard SORT with fixed R. Please provide a justification for this functional form, report the parameter values, and include a sensitivity analysis or a dedicated ablation so that the contribution of the adaptive component can be assessed.","section":"III-A, Eq. (4)"}],"minor_comments":[{"comment":"There are numerous typographical errors, including 'futher' (Section IV-B.3 and Conclusion), 'calulate' (Section III-A), 'their are' (Section III-A), 'ASDUGS-VINS' (Section IV-B.2), and a stray 'e' in the heading '1) ATE Comparison on VIODE Dataset: e'. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The ablation discussion in Section IV-B.2 refers to a 'baseline' without defining it clearly, while the caption of Fig. 6 states the comparison is 'including SORT (red) and excluding SORT (blue)'. Please clarify what constitutes the baseline in the heatmap ablation and how it relates to the SORT on/off comparison.","section":"IV-B.2 and Fig. 6"},{"comment":"The description of the OpenLORIS dataset says it contains 'visual, inertial, and odometric data,' but the experimental setup does not specify which sensor subset is used, how the VIO methods are initialized, or whether loop closure is enabled. Please provide this information, as it affects reproducibility.","section":"IV-A"},{"comment":"The values of the sliding-window length N, the adaptive Kalman parameters lambda and beta, and the feature parameters Nmax and Dmin used in the experiments are not reported. Please include them in the text or in a table so that the results can be reproduced.","section":"III-A"},{"comment":"The real-world experiment is only qualitative, showing trajectory overlays on satellite imagery without ground truth or quantitative error metrics. Please either provide a quantitative evaluation (e.g., with surveyed ground control points) or rephrase the claim to make clear that the real-world results are a demonstration rather than a benchmark.","section":"IV-D"}],"recommendation":"major_revision","confidential_remarks":"This paper addresses a relevant problem and the main idea is reasonable, but the current evidence is not yet strong enough for a clear accept. The two most important concerns are (1) the lack of any evaluation of the motion-agnostic masking policy, which is the core mechanism, and (2) the partially non-self-contained OpenLORIS comparison. If the authors can add a mask-quality analysis and re-run baselines, the paper could be suitable for publication. The adaptive Kalman component also needs a sensitivity analysis. The manuscript would benefit from professional proofreading."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"A credible engineering integration of existing pieces into VINS-Mono: YOLOv11 detections, an enhanced SORT tracker whose Kalman measurement noise R is adaptively scaled by an erf function of residual RMSE, and Mobile SAM promptable segmentation to mask dynamic objects. The new bits are the adaptive erf-based R update and the compensation strategy based on ANMS, plus a small real-world dataset. That is a legitimate incremental contribution to dynamic VIO.\n\nWhat it does well: the pipeline is described clearly enough to re-implement, the ablation shows the SORT component helps (83% ATE reduction on some VIODE sequences), and the VIODE comparison table is internally consistent. Releasing the real-world dataset is a nice gesture.\n\nSoft spots, in order. First, the masking is motion-agnostic: every detected car, pedestrian, or cart is masked regardless of whether it moves. Static objects of those classes get their features discarded, which can only hurt the optimizer. The paper claims generalization but gives no mask precision/recall, no count of static features wrongly removed, and no check of dynamic features left in. This is the load-bearing risk, and it is not addressed. Second, the empirical support for 'outperforms state-of-the-art' is thin: no error bars or repeated trials, some OpenLORIS baselines are copied from the prior OpenLORIS paper, the real-world evaluation is qualitative only, and the key tuning parameters lambda and beta are not reported. Code is not released either. These are addressable in revision.\n\nThe stress-test concern about static-feature discard is on point; the central pose claim is not circular, and the benchmark comparison is external, so this is a matter of missing evidence rather than a fundamental flaw.\n\nWho it is for: people working on dynamic VIO or semantic masking in visual-inertial systems. A serious referee should engage, but would ask for the missing metrics and code before accepting.\n\nRecommendation: send to peer review, with the expectation of a major revision that adds mask-quality evaluation, repeated runs with error bars, and parameter disclosure.","headline":"A plausible dynamic-VIO integration of YOLOv11, adaptive SORT, and Mobile SAM that needs better empirical support before the generalization claim holds.","tokens_in":12201,"tokens_out":1849,"would_cite":true,"duration_ms":16609,"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":"The paper claims that a track-then-segment front end that masks moving objects and discards the feature points inside them substantially improves visual-inertial odometry accuracy in highly dynamic, partially occluded scenes.","keywords":["visual-inertial odometry","dynamic environments","dynamic object segmentation","Segment Anything Model","SORT tracking","adaptive Kalman filter","feature point rejection","absolute trajectory error"],"falsifier":"Take the released real-world dataset, hand-label every moving object in every frame, and compare the masks produced by the pipeline against those labels; if frames where the mask misses a moving object show no increase in per-frame pose error, the central claim that mask-based rejection produces the gains is false.","tokens_in":11124,"feed_emoji":"🧭","tokens_out":8492,"duration_ms":72713,"temperature":0.7,"pith_summary":"The paper claims that visual-inertial odometry can be made substantially more accurate in dynamic scenes by running a track-then-segment front end: track every potentially moving object, cut it out with a promptable foundation model that turns a box prompt into a pixel mask, and discard any image feature points that fall inside those masks before optimization. The payoff, if the claim holds, is a single VIO pipeline that handles many kinds of moving objects, such as cars, pedestrians, and shopping carts, without retraining the segmenter for each scene. The paper reports lower absolute trajectory error than several established VIO systems on synthetic and real indoor and outdoor datasets, with the largest gains in sequences where dynamic objects dominate the view. The paper also releases a real-world visual-inertial dataset with diverse moving characters so the dynamic-object setting can be benchmarked directly.","feed_headline":"Moving-object masks cut VIO trajectory error in dynamic scenes","feed_subtitle":"Track-then-segment removes moving points before bundle adjustment, with reported gains up to 83 percent on one dataset.","key_machinery":"The load-bearing mechanism is the dynamic-object mask generated from the tracked boxes, because that mask decides which feature points are excluded from visual-inertial bundle adjustment. Tracking is stabilized by an adaptive Kalman update in which the measurement noise covariance is $\\mathbf{R} = \\mathrm{diag}(\\beta\\, \\mathrm{erf}(\\lambda\\, \\delta_{\\mathrm{RMSE}}))$, so the filter becomes more cautious when residual errors grow. The boxes prompt Mobile SAM to segment the object at pixel level, and erosion followed by a slightly larger dilation removes mask noise while pushing the mask beyond object edges so no boundary features survive. That mask, not geometric outlier rejection, is what carries the dynamic-feature removal in the pipeline.","core_discovery":"On the paper's own terms, the core discovery is that a track-then-segment-then-discard pipeline plugged into a standard optimization-based VIO front end produces pose estimates that beat the compared systems in dynamic scenes. The pipeline detects candidate moving objects, associates them across frames with an adaptive Kalman filter whose measurement-noise covariance is scaled by an error function of the sliding-window residual RMSE, and uses the tracked boxes to prompt Mobile SAM, a lightweight variant of the Segment Anything foundation model, for pixel-accurate masks. KLT-tracked feature points inside the masks are removed; ORB points are re-extracted in unmasked regions under adaptive non-maximal suppression, and a compensation rule keeps the static feature count near the maximum. The paper reports lower RMSE absolute trajectory error on the VIODE high-dynamic sequences, stable tracking on OpenLORIS market scenes, and less drift than the baseline on a real-world outdoor route.","pith_inferences":["The same mask-and-discard module could likely be grafted onto other VIO or SLAM systems, so the main transferable result is the pipeline rather than the specific detector, tracker, or segmenter.","Because the published numbers never measure mask precision or recall, an immediate test is to correlate per-frame mask quality against per-frame pose error on the released dataset.","A harder stress test the paper does not report is an object class the detector has never seen or an occlusion that breaks the tracker's data association; those failures would reveal how much of the universality comes from the foundation model versus the detector and tracker.","The paper lists real-time performance as future work, so the practical claim is accuracy, not yet deployment at interactive rates."],"forward_implications":["Dynamic-object masking can be layered onto existing optimization-based VIO front ends, since the reported gains come without changing the IMU or visual bundle-adjustment core.","Sequences that previously caused VIO failure, such as the asterisked ORB-SLAM3 cases in the VIODE table, become tractable because the rejected features no longer pull the optimization.","The zero-shot segmenter should let the system move between outdoor traffic and indoor crowds without retraining, as demonstrated by the transfer from VIODE to OpenLORIS.","The SORT ablation isolates a large share of the improvement in the tracker: a 21.38% ATE reduction in the parking lot, 83.28% in city day, and 63.99% in city night.","A publicly released real-world dataset with diverse moving characters gives other methods a common benchmark for dynamic VIO."],"supporting_citations":[{"why":"Supplies the promptable foundation model whose masks mark dynamic regions for feature rejection.","marker":"[6]"},{"why":"Provides the base SORT tracking pipeline that the adaptive Kalman filter extends.","marker":"[7]"},{"why":"YOLOv11 detects candidate dynamic objects and supplies the boxes that prompt the segmentation model.","marker":"[30]"},{"why":"KLT sparse optical flow tracks feature points whose mask membership is checked each frame.","marker":"[33]"},{"why":"ORB detects new feature points in the unmasked regions after dynamic points are removed.","marker":"[34]"},{"why":"Dyna-VINS is the closest semantic dynamic-VIO baseline and the prior approach the paper's design builds past.","marker":"[3]"},{"why":"ORB-SLAM3 is a non-semantic VIO baseline whose failures in the VIODE table motivate the masking approach.","marker":"[8]"},{"why":"VINS-Mono is a standard monocular VIO baseline used in the VIODE comparison and in the real-world trajectory test.","marker":"[17]"},{"why":"VIODE supplies the synthetic outdoor dataset with controlled dynamic levels used for the main ATE comparison and ablations.","marker":"[36]"},{"why":"OpenLORIS supplies the indoor scenes and correct-rate metric used to argue generalization beyond traffic scenes.","marker":"[37]"}],"fun_headline_variants":["Track-then-segment VIO beats dynamic-scene baselines","ADUGS-VINS: masks moving objects to cut VIO drift","Slice dynamic objects out of VIO with promptable masks","Moving-point removal sharpens VIO in clutter","Foundation-model masks boost dynamic-scene VIO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole accuracy gain rests on the detector, tracker, and segmenter catching every moving object and masking exactly the dynamic pixels, with no missed moving objects and no wrongly masked static regions.","fun_headline_variants_meta":{"raw":{"variants":["Track-then-segment VIO beats dynamic-scene baselines","ADUGS-VINS: masks moving objects to cut VIO drift","Slice dynamic objects out of VIO with promptable masks","Moving-point removal sharpens VIO in clutter","Foundation-model masks boost dynamic-scene VIO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1244,"prompt_tokens":896,"completion_tokens":348,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":266}},"tokens_in":512,"tokens_out":348,"duration_ms":3737,"temperature":1.0,"reasoning_tokens":266,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:19:15.927081+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the released real-world dataset, hand-label every moving object in every frame, and compare the masks produced by the pipeline against those labels; if frames where the mask misses a moving object show no increase in per-frame pose error, the central claim that mask-based rejection produces the gains is false.","supporting_citations":[{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"Supplies the promptable foundation model whose masks mark dynamic regions for feature rejection."},{"cited_title":"Orb: An efficient alternative to sift or surf,","cited_arxiv_id":null,"evidence_quote":"ORB detects new feature points in the unmasked regions after dynamic points are removed."},{"cited_title":"Dynavins: A visual- inertial slam for dynamic environments,","cited_arxiv_id":null,"evidence_quote":"Dyna-VINS is the closest semantic dynamic-VIO baseline and the prior approach the paper's design builds past."},{"cited_title":"Viode: A simulated dataset to address the challenges of visual-inertial odometry in dynamic environments,","cited_arxiv_id":null,"evidence_quote":"VIODE supplies the synthetic outdoor dataset with controlled dynamic levels used for the main ATE comparison and ablations."},{"cited_title":"Are we ready for service robots? the openloris-scene datasets for lifelong slam,","cited_arxiv_id":null,"evidence_quote":"OpenLORIS supplies the indoor scenes and correct-rate metric used to argue generalization beyond traffic scenes."}],"review_version":1}