{"id":"57ad9928-3fa1-4cf0-bca2-b7e98f6cdbc2","arxiv_id":"2501.13183","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"MONA detects moving objects by filtering YOLO boxes with LEAP-VO dynamic point probabilities and optical flow, then segments them with SAM; integrated with LEAP-VO it reports state-of-the-art trajectory accuracy on MPI Sintel.","lead":"MONA is a pipeline that detects and segments moving objects in videos from moving cameras by combining point tracking, optical flow, object detection, and a segmentation model. Tested as a front-end for the LEAP-VO visual odometry system, it reports large accuracy gains on the synthetic MPI Sintel benchmark.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncontrolled proxy: VO gains are not shown to come from moving-object detection rather than generic point filtering; direct ablation missing.","rationale":"The reader's verdict of REJECT is well-founded. The paper's central claim is that MONA performs moving object detection and segmentation accurately, but the only quantitative support is the downstream camera trajectory improvement on MPI Sintel. This is an indirect proxy, and the paper itself acknowledges that no direct evaluation of detection or mask quality is performed. The most load-bearing weakness is the absence of a control experiment: the improvement could stem from generic point filtering rather than the specific dynamic/static distinction that MONA is designed to make. The reader identified exactly this assumption, and I agree. In addition, the missing formula for the adaptive threshold θt and the vague integration description with LEAP-VO make the method irreproducible and prevent independent verification of the mechanism. The reported 'over 60% improvement' is thus not convincingly attributed to the paper's claimed contribution. A single concrete experiment—comparing MONA against simpler point-exclusion baselines—would settle whether the causal claim holds, but as written the central claim is unsupported.","tokens_in":7717,"tokens_out":5912,"duration_ms":55108,"concrete_test":"Run the following controlled experiment on the same MPI Sintel sequences and report ATE/RPE: (1) LEAP-VO with MONA as described; (2) LEAP-VO with all YOLO-detected boxes (static and dynamic) used as the exclusion mask; (3) LEAP-VO with points excluded solely on optical-flow magnitude (e.g., flow > c·mean, c∈{1,2,3}) instead of MONA's dynamic-point threshold. Keep every other setting identical. If (2) or (3) matches (1) within 10% on all three metrics, the claimed causal role of moving-object detection and segmentation is not established. Additionally, provide the exact formula for θt and verify that the adaptive threshold outperforms a fixed threshold.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim is that MONA's moving object detection and segmentation quality drives a >60% improvement in camera trajectory estimation on MPI Sintel (Table 1). The load-bearing assumption is that the improvement is specifically caused by identifying and masking moving objects, not by a generic filtering effect. This assumption is untested. The paper itself states 'no public dataset is available to directly evaluate detection accuracy and mask quality,' and the only quantitative evidence is the downstream ATE/RPE table. The integration with LEAP-VO is described only at a high level ('the randomly selected points inside the moving objects mask can be filtered in the bundle adjustment process'), with no pseudocode, no detail on how many points are removed, and no specification of the adaptive threshold θt (Eq. 5 only defines the mean flow magnitude, not θt). Crucially, there is no ablation replacing MONA's masks with a simpler filter—e.g., removing all points in any YOLO box, or all points whose optical flow exceeds a multiple of the mean—to show that the gains require the dynamic/static distinction. Without such a control, the observed improvement could be due to generic outlier rejection or dataset-specific point selection, and the central claim that MONA achieves accurate moving object detection is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MONA, a two-module framework for detecting and segmenting moving objects in videos captured by dynamic cameras. The first module extracts dynamic points by combining dynamic probabilities from LEAP-VO with RAFT optical flow and an adaptive threshold; the second module filters YOLO bounding boxes using those dynamic points and feeds them as prompts to SAM to obtain segmentation masks. The method is validated indirectly by integrating it with LEAP-VO for camera trajectory estimation on the MPI Sintel dataset, where Table 1 reports ATE 0.029 m, RPE trans 0.013 m, and RPE rot 0.054 deg, described as over 60% improvement over LEAP-VO. No direct metric for moving object detection boxes or segmentation masks is reported.","tokens_in":52,"tokens_out":3914,"duration_ms":98069,"significance":"If MONA genuinely improves moving-object detection and segmentation in dynamic-camera videos, the contribution would be useful for visual odometry and downstream urban-scene applications. The integration design is reasonable: reusing an established dynamic-probability model, combining off-the-shelf components (RAFT, YOLO, SAM), and demonstrating a downstream trajectory improvement is a sensible way to show utility. I also see no circularity: MONA is not trained to optimize the reported ATE/RPE numbers, and LEAP-VO serves both as a component source and as an evaluation harness. However, the significance of the claimed detection capability rests entirely on an indirect proxy, and the paper currently does not establish that the downstream gain is caused by accurate moving-object segmentation rather than by generic point filtering or dataset-specific effects.","major_comments":[{"comment":"The first paragraph of Experiments and Results states that no public dataset is available to directly evaluate detection accuracy and mask quality, and the only quantitative evidence presented is Table 1, which reports ATE and RPE from camera trajectory estimation. This is a load-bearing gap: the paper's title and abstract claim a moving object detection and segmentation capability, but no metric on detected boxes or predicted masks is reported. The downstream proxy is only informative if the causal link between masking quality and trajectory accuracy is established, which the paper does not do. Direct evaluation on a dataset with moving-object masks (e.g., using the MPI Sintel ground-truth segmentation or a synthetic moving-camera set) or at least mask/box metrics on a suitable benchmark is needed to support the central claim.","section":"Experiments and Results"},{"comment":"The claimed >60% improvement over LEAP-VO is attributed to MONA's moving-object detection and segmentation, but no control experiment isolates this mechanism. An ablation replacing MONA's masks with simpler filters, such as removing all points whose optical flow magnitude exceeds a multiple of the mean, removing points inside all YOLO boxes, or randomly removing a matched number of points, would test whether the gain requires the dynamic/static distinction. The existing ablation in Fig. 3 only compares segmentation mask appearance for three prompt types; it does not quantify the effect of these masks on trajectory estimation. Without such a control, the improvement in Table 1 could be due to generic outlier rejection or dataset-specific point selection rather than accurate moving-object segmentation.","section":"Table 1 and Ablation Study"},{"comment":"The adaptive threshold θt is never defined. Eq. (5) defines the mean optical flow magnitude m̄_t, and the text says the threshold is 'dynamically scaled' from m̄_t, but the scaling rule is not specified. Similarly, Eq. (6) introduces τ0 as a threshold for dynamic points inside a bounding box, but the paper does not state its value, how it is selected, or how sensitive results are to it. The parameters n, m, k (number of detection points, anchor points, grid size) and λ in Eq. (4) are also left unspecified. These free parameters are load-bearing for reproducibility and for assessing whether the thresholding is genuinely adaptive, so they need to be concretely defined.","section":"Methods, Eq. (5) and Eq. (6)"},{"comment":"Table 1 reports single-point estimates without error bars, confidence intervals, or repeated-run statistics. The pipeline includes stochastic components (random selection of detection points, CoTracker, RAFT, YOLO, SAM, and random point filtering in LEAP-VO), so run-to-run variability is likely non-negligible. The claim of state-of-the-art performance and the quantitative 'over 60% improvement' should be supported by variance information or multiple runs.","section":"Table 1"}],"minor_comments":[{"comment":"Eq. (2) uses p(x|V, xq) on both sides of the product; the second factor should be p(y|V, xq), and the product structure should be stated clearly to avoid confusion between the coordinate symbol and the conditioning variables.","section":"Methods, Eq. (2)"},{"comment":"There is a typo in 'UA V' (should be 'UAV'), and the phrase 'In Dynamic Points Extraction , random points are selected' contains an extra space before the comma. A final proofreading pass is needed.","section":"Introduction"},{"comment":"The qualitative trajectory comparison in Fig. 2 would be more informative with quantitative per-sequence results and a description of how many Sintel sequences show improvement or regression. A per-sequence breakdown of ATE/RPE would strengthen the claim beyond aggregate numbers.","section":"Experiments and Results"},{"comment":"Several references in the introduction, such as many smart-city and urban-planning citations, are only loosely connected to the technical content. The authors should consider focusing the related-work discussion on moving object detection, dynamic SLAM/VO, and segmentation methods.","section":"References"},{"comment":"The notation u= arg min_{b in B} {Area(b) | D_u ≥ τ0} is ambiguous because the subscript u appears in the condition and in the set being minimized; this should be rewritten for clarity, for example by defining the set of candidate boxes first.","section":"Moving Object Segmentation, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central gap is not a mathematical inconsistency but an unsupported causal attribution: the reported trajectory improvements are not tied to measured detection or segmentation quality. I believe the gap can be addressed within the manuscript's scope by adding direct mask/box evaluation on a suitable benchmark and by ablating the moving-object mask against generic point-filtering baselines. If the authors can provide such evidence, the contribution would be substantially stronger; as presented, the evidence is too indirect for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about MONA is that it assembles a clean, sensible pipeline for masking moving objects—CoTracker/LEAP-VO dynamic probabilities, RAFT flow, YOLO boxes, an area-scaled filter, and SAM prompts—but the paper never directly measures the quality of the boxes or masks it produces. The only quantitative evidence is a downstream camera-trajectory table on MPI Sintel, and that table cannot carry the paper's stated claim.\n\nWhat is genuinely new is the combination. I haven't seen this exact chain in the literature, and the adaptive bounding-box filter (Eq. 6) is a reasonable heuristic that addresses the static/dynamic box imbalance. The qualitative ablation in Fig. 3 shows the filtered boxes give cleaner SAM masks than raw dynamic points or unfiltered YOLO boxes. Credit where it's due: the authors are upfront in the experiments section that no public dataset exists for direct evaluation of moving-object detection in dynamic-camera video, and they choose the downstream task deliberately.\n\nThe soft spots are real and load-bearing. First, the causal claim is untested. The >60% improvement on ATE/RPE could come from generic outlier rejection—removing any points in a YOLO box, or any high-flow points—rather than from correctly separating moving from static objects. No such control is run. Second, θt is never actually defined: Eq. 5 computes the mean optical flow magnitude, but the scaling rule that turns it into θt is missing. Third, τ0 is left as a free parameter with no sensitivity analysis, and Table 1 has no error bars. The conclusion that 'these results validate the effectiveness of our approach' overreaches; at best they validate MONA+LEAP-VO as a combined trajectory estimator.\n\nThe result itself may be useful. If the system genuinely gives a 60% trajectory improvement on Sintel, that is practically interesting for dynamic-scene VO/SLAM, and the integration is worth knowing about. But as a paper about moving object detection and segmentation, the current evidence doesn't support the title claim.\n\nI'd send this to peer review—it deserves a serious referee who can demand direct metrics (e.g., on KITTI or DAVIS) and the missing generic-filter ablation. Without those, it should not be accepted. My own verdict would be reject in present form, but it's a close call and the idea has merit.","headline":"Clean new pipeline for masking moving objects, but the evaluation never measures detection quality directly and the trajectory gains are an uncontrolled proxy.","tokens_in":8493,"tokens_out":2707,"would_cite":false,"duration_ms":27931,"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":"A moving-object detection and segmentation pipeline for dynamic-camera video, integrated with LEAP-VO, cuts trajectory errors by more than 60% on MPI Sintel.","keywords":["moving object detection","dynamic camera","optical flow","point tracking","video segmentation","visual odometry","camera trajectory estimation","MPI Sintel"],"falsifier":"Run MONA on MPI Sintel while replacing its moving-object masks with masks of equal area placed at random, or with a fixed fraction of low-confidence tracked points removed; if trajectory error stays near MONA+LEAP-VO's 0.029 m ATE, the segmentation accuracy is not the cause. Alternatively, compute the IoU between MONA's masks and hand-annotated moving-object masks on any dynamic-camera video; low IoU alongside large trajectory gains would show the claimed mechanism is not what drives the improvement.","tokens_in":7554,"feed_emoji":"🎥","tokens_out":5655,"duration_ms":51137,"temperature":0.7,"pith_summary":"The paper introduces a two-module pipeline, MONA, for detecting and segmenting moving objects in videos recorded by moving cameras. Its first module scores how likely tracked points are to be moving, then uses optical-flow magnitudes to set a per-frame threshold that separates dynamic from static points. The second module feeds those dynamic points into an object detector, filters candidate bounding boxes by how many dynamic points they contain, and uses the surviving boxes as prompts for a segmentation model. To show the detection quality matters, the authors integrate MONA with the visual odometry method LEAP-VO and report state-of-the-art trajectory accuracy on the MPI Sintel dataset, with over 60% improvement in absolute and relative trajectory errors over LEAP-VO alone. The paper's central assertion is that accurate moving-object masks, not generic filtering, drive this downstream gain.","feed_headline":"Masking moving objects cuts camera trajectory error 60% on Sintel","feed_subtitle":"A dynamic-camera pipeline detects and masks moving objects, then feeds the masks to visual odometry to sharpen trajectory estimates.","key_machinery":"The central mechanism is the adaptive bounding-box filter with an area-scaled threshold: for each YOLO box $b_i$, the dynamic-point count $D_i$ is compared against $\\tau_t^i = \\tau_0 \\times \\mathrm{Area}(b_i)/\\mathrm{Area}(b_u)$, where $b_u$ is the smallest box whose count exceeds the base threshold $\\tau_0$. A second mechanism is the frame-adaptive dynamic-point threshold: the mean optical-flow magnitude $\\bar{m}_t$ of a frame scales the threshold $\\theta_t$ that classifies points as dynamic. These two scaling rules let the pipeline work without a fixed hyperparameter across frames with different camera motion and object sizes.","core_discovery":"MONA claims that moving objects in dynamic-camera video can be identified without a static background model. The dynamic-points module reuses LEAP-VO's anchor-based probability estimate that each queried pixel is part of a moving object, then refines it with RAFT's optical flow: the mean flow magnitude of a frame sets the threshold for declaring points dynamic. The segmentation module detects all objects with YOLO, keeps only bounding boxes whose dynamic-point count exceeds an area-scaled threshold, and prompts SAM with those boxes to produce object masks. Integrated into LEAP-VO's bundle adjustment, the masks filter out tracked points inside moving regions, yielding an ATE of 0.029 m, relative translation error of 0.013 m, and relative rotation error of 0.054 degrees on MPI Sintel, state-of-the-art among the compared methods. The paper interprets these numbers as evidence that MONA's detection and segmentation are accurate enough to improve downstream camera trajectory estimation.","pith_inferences":["The reported trajectory gains could come from generic removal of outlier tracks rather than from semantically accurate masks; a control that filters the same number of points at random would separate these explanations.","If the thresholding rules are the real driver, MONA may generalize to moving cameras and objects beyond the categories YOLO is trained on, since no fine-tuning is required.","One testable extension is to report detection and mask metrics such as IoU and recall on a dynamic-camera benchmark; the current evidence for detection quality is indirect, through trajectory accuracy.","The adaptive area-scaling threshold could be reused as a cheap prior for other prompt-based segmentation tasks where object size varies."],"forward_implications":["Masking dynamically moving regions during bundle adjustment lets LEAP-VO select tracked points almost entirely from static scene structure, which is why trajectory errors drop by more than 60 percent on MPI Sintel.","The same masked frames can be fed to other visual odometry and SLAM systems, potentially improving their accuracy in urban scenes with large pedestrians or vehicles.","Reliable moving-object masks make markerless online video usable for generating pseudo-ground-truth camera trajectories for autonomous driving, UAV planning, and human motion recovery.","Because MONA needs no training on moving-object annotations, it can be applied to new domains where dynamic objects appear but labeled video is unavailable."],"supporting_citations":[{"why":"Supplies LEAP-VO, the anchor-based dynamic probability estimation that MONA reuses and the downstream trajectory estimator that MONA augments.","marker":"Chen et al. 2024b"},{"why":"Supplies CoTracker, the tracking-any-point method that produces the trajectories and visibility used to compute point dynamic probabilities.","marker":"Karaev et al. 2024"},{"why":"Supplies RAFT, the optical flow method whose mean magnitude sets the frame-adaptive threshold for classifying points as dynamic.","marker":"Teed and Deng 2020"},{"why":"Supplies YOLO, the object detector whose bounding boxes are filtered by dynamic-point counts to identify moving objects.","marker":"Khanam and Hussain 2024"},{"why":"Supplies the Segment Anything Model (SAM), which turns the filtered bounding boxes into precise moving-object masks.","marker":"Kirillov et al. 2023"},{"why":"Provides the MPI Sintel dataset, including ground-truth camera trajectories, used for the quantitative evaluation.","marker":"Butler et al. 2012"}],"fun_headline_variants":["MONA detects moving objects in dynamic-camera video without static model","Optical flow + SAM in MONA: SOTA on MPI Sintel for dynamic cameras","Dynamic camera footage? MONA masks moving objects to boost odometry accuracy","MONA sets new state-of-the-art on Sintel for moving-object detection","MONA's pipeline: dynamic points, SAM masks, better LEAP-VO trajectory"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim collapses if the trajectory improvements on MPI Sintel come from generic filtering of unreliable tracked points rather than from MONA's moving-object detection and segmentation, because the paper reports no direct metrics for detection or mask quality.","fun_headline_variants_meta":{"raw":{"variants":["MONA detects moving objects in dynamic-camera video without static model","Optical flow + SAM in MONA: SOTA on MPI Sintel for dynamic cameras","Dynamic camera footage? MONA masks moving objects to boost odometry accuracy","MONA sets new state-of-the-art on Sintel for moving-object detection","MONA's pipeline: dynamic points, SAM masks, better LEAP-VO trajectory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000449,"raw_usage":{"total_tokens":2230,"prompt_tokens":874,"completion_tokens":1356,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":1253}},"tokens_in":490,"tokens_out":1356,"duration_ms":13925,"temperature":1.0,"reasoning_tokens":1253,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:22:32.912217+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MONA on MPI Sintel while replacing its moving-object masks with masks of equal area placed at random, or with a fixed fraction of low-confidence tracked points removed; if trajectory error stays near MONA+LEAP-VO's 0.029 m ATE, the segmentation accuracy is not the cause. Alternatively, compute the IoU between MONA's masks and hand-annotated moving-object masks on any dynamic-camera video; low IoU alongside large trajectory gains would show the claimed mechanism is not what drives the improvement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies CoTracker, the tracking-any-point method that produces the trajectories and visibility used to compute point dynamic probabilities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies RAFT, the optical flow method whose mean magnitude sets the frame-adaptive threshold for classifying points as dynamic."},{"cited_title":"J.; Wulff, J.; Stanley, G","cited_arxiv_id":null,"evidence_quote":"Provides the MPI Sintel dataset, including ground-truth camera trajectories, used for the quantitative evaluation."}],"review_version":1}