{"id":"c5985e4f-b0ae-4258-90a9-83e908cde51b","arxiv_id":"2507.01455","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"OoDDINO combines an uncertainty-fused open-set detector with region-adaptive dual thresholds to improve road-scene anomaly segmentation over RPL and RbA baselines.","lead":"This paper presents OoDDINO, a two-stage system that first uses an open-set object detector boosted by multiple uncertainty signals to find anomalous objects in road images, then applies different score thresholds inside and outside those detected regions to segment anomalies. If the reported gains hold, the method offers a practical way to reduce false alarms in autonomous driving perception without retraining existing pixel-wise anomaly models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ADT-Net's normalization range α is tuned on the test set, and main/ablation tables report conflicting numbers for identical configurations, so the reported improvements are not trustworthy.","rationale":"The reader's weakest_assumption focused on reliance on object-detector quality, which is a legitimate design concern but is not the most damaging issue. The reader's rationale, however, did note that the normalized score interval is selected using numbers that match the final headline results, which is the test-set tuning I identify as the load-bearing flaw. I partially agree because the reader did not make this the central concern. Test-set tuning invalidates the statistical validity of all reported benchmark numbers: the claimed gains from 17.74% to 4.78% FPR95 and from 71.6% to 87.1% AP may simply reflect the choice of α that maximized performance on those same test images. This is a fundamental methodological error that prevents acceptance of the current evidence, regardless of the architectural soundness. The table inconsistency compounds the issue by making it impossible for a reader to know which numbers are reliable. Therefore, the paper should be rejected as submitted; it could be reconsidered if the authors re-run experiments with validation-based hyperparameter selection and provide a reconciled set of results.","tokens_in":151,"tokens_out":4584,"duration_ms":59550,"concrete_test":"Re-run the full OoDDINO pipeline on SMIYC and RoadAnomaly, selecting α (and any other tuned hyperparameters) using a validation split (e.g., a held-out subset of the synthetic training data or a validation set from the target datasets) before computing test metrics. Report test FPR95/AP for α ∈ {0.2, 0.3, 0.4} chosen by validation, not by test performance. Also provide a version of Tables 1 and 3 that uses the same evaluation protocol so the numbers match. If validation-selected α yields materially worse results than reported (e.g., RoadAnomaly AP drops more than 5 points or FPR95 doubles), the claim of large improvements is an artifact of test-set tuning.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim—that OoDDINO substantially reduces FPR95 and improves AP over frozen baselines—rests on the empirical results in Tables 1–3. Two issues undermine this evidence. First, Appendix C explicitly states that the normalization-range parameter α in Eq. (2) was set by varying its value from 0.1 to 0.4 and selecting the range [0.3, 0.8] that gave the best FPR/AP on the RoadAnomaly test set. This is test-set hyperparameter tuning: the reported AuROC of 98.9% and FPR95 of 4.78% are not unbiased estimates but the best of several configurations evaluated on the final test set. A plug-in that appears to boost performance by 13 points of AP may simply be overfit to these benchmarks. Second, the same configurations are inconsistent across tables: Table 1 lists RPL+Ours on AnomalyTrack with AP 87.3 and FPR95 7.8, while Table 3 lists OoDDINO on AnomalyTrack with AP 92.08 and FPR95 3.82; the RPL baseline also differs (83.4/11.7 vs 88.55/7.18). If the ablation numbers are correct, the relative gains are different and the absolute improvements are larger, but if Table 1 is correct, Table 3 overstates the effect. Either way, the reported gains cannot be independently verified from the paper as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OoDDINO, a two-stage framework for anomaly segmentation in road scenes. In the first stage, an Orthogonal Uncertainty-Aware Fusion Strategy (OUAFS) integrates semantic-segmentation, softmax-entropy, and softmax-distance maps into a GroundingDINO open-set detector under an orthogonal-loss constraint. In the second stage, an Adaptive Dual-Threshold Network (ADT-Net) normalizes pixel-wise anomaly scores and learns separate foreground/background thresholds conditioned on detection proposals. The anomaly segmentation baselines RPL and RbA are kept frozen, and the paper reports substantial gains in FPR95, AP, and component-level F1 on SMIYC, RoadAnomaly, and Fishyscapes, positioning OoDDINO as a plug-in module for existing pixel-wise anomaly segmentation methods.","tokens_in":16777,"tokens_out":4263,"duration_ms":46942,"significance":"If the reported results are correct, the work is a useful and fairly original contribution: it is, to my knowledge, the first systematic use of an open-set detector as an object-level prior for anomaly segmentation, and the proposed two-threshold mechanism directly addresses a known weakness of global thresholding. The modular design is a genuine strength, as is the release of source code. However, the empirical evidence for the central claim currently has two load-bearing problems: the same configuration is reported with different numbers in Tables 1 and 3, and the normalization-range parameter alpha in Eq. (2) is selected on the test set in Appendix C. These issues must be resolved before the headline improvements can be trusted.","major_comments":[{"comment":"The same RPL-based configuration is reported with inconsistent numbers on AnomalyTrack: Table 1 gives RPL an AP of 83.4 and FPR95 of 11.7, while Table 3 gives RPL an AP of 88.55 and FPR95 of 7.18; similarly, Table 1 reports RPL+Ours with AP 87.3 and FPR95 7.8, while Table 3 reports OoDDINO with AP 92.08 and FPR95 3.82. Because Tables 1 and 3 are supposed to evaluate the same framework on the same benchmark, this discrepancy changes the claimed improvement by roughly 5 AP points and 4 FPR95 points. Please reconcile the two tables and report a single consistent evaluation protocol, or explain explicitly why the numbers differ.","section":"Tables 1 and 3"},{"comment":"The normalization-range parameter alpha is tuned on the RoadAnomaly test set: Appendix C states that the lower bound was varied from 0.1 to 0.4 and that [0.3, 0.8] was selected because it gave the best FPR/AP. As a result, the headline numbers, including the 98.9% AuROC and the 4.78% FPR95, are the best of several test-set evaluations rather than estimates under a fixed protocol. This makes the reported gains look potentially overfit to the test benchmarks. Please fix alpha using a validation split or a pre-registered default, and report test-set performance for the single chosen value.","section":"Appendix C, Eq. (2)"},{"comment":"The foreground mask is defined as Mfg = 1 in {1}^{H x W} and the background mask as Mbg = 1 - Mfg, which makes the background branch empty and contradicts the surrounding text, which says the masks are derived from detection proposals. As written, Eq. (2) does not specify the advertised region-adaptive mechanism. Please correct the definition to Mfg = 1_{boxes} (or an equivalent indicator over detection proposals) and define Mbg accordingly.","section":"Sec. 3.3, Eq. (2)"},{"comment":"All experiments are reported as point estimates without error bars, multiple seeds, or confidence intervals, on small test sets (RoadAnomaly has 60 images and AnomalyTrack has 100). Many of the claimed improvements are only a few AP or FPR95 points, so without repeated-run statistics it is unclear whether the differences are within run-to-run noise. Please add standard deviations over at least three runs, or otherwise justify the stability of the reported numbers.","section":"Sec. 4.1, Tables 1-3"},{"comment":"The ablation shows that naively adding GroundingDINO to RPL degrades all metrics, e.g., RoadAnomaly FPR95 rises from 17.74 to 28.50. The final framework therefore depends critically on OUAFS and ADT-Net correcting detector errors, but the paper does not analyze cases where the detector still misses anomalies or outputs spurious boxes, nor does it report detector recall on the benchmarks. Please add a quantitative analysis of detection failures and a discussion of how the framework behaves when the object-level proposals are unreliable.","section":"Sec. 4.3, Table 3"}],"minor_comments":[{"comment":"There are typos in the text: 'segemntation' in Section 4.1 and 'ObstleTrack' in Section 4.2.1 should be corrected.","section":"Sec. 4.1, 4.2.1"},{"comment":"References [9] and [10] appear to be the same paper, and [19] and [20] are also duplicates; please merge them and renumber.","section":"References"},{"comment":"The caption says 'the two leftmost columns display the image and its ground truth,' but the described figure and surrounding text imply more columns; please make the caption consistent with the actual figure layout.","section":"Figure 3"},{"comment":"The Small/Medium/Large columns in Table 4 are not defined in the main text; please state what these categories refer to (e.g., object sizes or bounding-box scales) and how they are computed.","section":"Table 4"},{"comment":"The abstract says experiments use 'two benchmark datasets,' but the paper reports results on SMIYC, RoadAnomaly, and Fishyscapes; please correct the count.","section":"Abstract and Sec. 4"},{"comment":"The text does not specify which model produces the semantic segmentation map S_i used in OUAFS, nor which text prompts are given to GroundingDINO for anomaly detection; please provide these implementation details.","section":"Sec. 3.2"},{"comment":"The claim that this is 'the first time an open-set detection model has been applied to anomaly segmentation' is strong and should be either verified against prior work or softened.","section":"Sec. 2.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be in camera-ready form, including a placeholder author line 'Trovato et al.' and an ACM reference format, which is unusual for a submission under review. More substantively, the discrepancy between Tables 1 and 3 is the single most important issue for the paper's credibility; if it cannot be resolved, the paper would not be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the OoDDINO manuscript. The short version: the architecture is coherent and the plug-in idea is genuinely useful, but the empirical section has two problems that need fixing before any of the headline numbers can be believed.\n\nWhat's new: the paper identifies two real failures of pixel-wise anomaly segmentation—fragmented predictions and the global-threshold trade-off—and addresses them with a two-stage cascade. OUAFS sequentially fuses uncertainty maps into an open-set detector, and ADT-Net learns separate thresholds for foreground and background. That's a sensible design. The ablations in Table 3 show each module contributes, and Appendix D extends the framework to PEBAL and Mask2Anomaly, which supports the compatibility claim. The qualitative results look plausible.\n\nThe soft spots are in the evaluation. Appendix C is the big one: the normalization lower bound α is chosen by running the model on the RoadAnomaly test set with several values and picking the one that gives the best FPR/AP. That is test-set tuning. The reported 98.9% AuROC and 4.78% FPR are the best of a small grid search, not unbiased estimates. The authors need to move that selection to a validation split or report the full grid and acknowledge the selection.\n\nWorse, Table 1 and Table 3 report incompatible numbers for the same configurations. RPL on AnomalyTrack shows AP 83.4 / FPR 11.7 in Table 1, but AP 88.55 / FPR 7.18 in Table 3. The OoDDINO configuration is AP 87.3 / 7.8 in Table 1 and AP 92.08 / 3.82 in Table 3. A discrepancy of 4–5 AP points and 4 FPR points is not noise; a referee cannot tell which table is correct, and the claimed relative gains change depending on which you read. The paper also lacks error bars, which matters on datasets of 60–100 images.\n\nThere's also an underspecified input: the three uncertainty maps (segmentation, entropy, distance)—where do they come from? They are presumably computed from the frozen baseline model, but the paper never says so explicitly.\n\nThis is for researchers working on road-scene anomaly segmentation, especially those building plug-in components. With a cleaned evaluation it could be a useful contribution, but as written the evidence doesn't back the headline numbers. I would not desk-reject it; the architecture is coherent enough that a serious referee could guide it to a solid paper. My recommendation: send to peer review with major revision—reconcile the tables, move α off the test set, add error bars, specify the uncertainty-map sources, and make the code (they already promise a GitHub link) actually usable.","headline":"Architecture is coherent and the plug-in idea is useful, but test-set-tuned hyperparameters and contradictory tables make the reported gains unverifiable.","tokens_in":17315,"tokens_out":3419,"would_cite":false,"duration_ms":33486,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","68U10"],"pacs":[],"model":"deepseek-v4-flash","headline":"OoDDINO claims that bolting an object-level anomaly detector onto frozen pixel-wise segmenters, then cutting the score map with two region-specific thresholds instead of one global threshold, sharply reduces false positives and raises AP…","keywords":["anomaly segmentation","open-set object detection","adaptive dual-threshold network","orthogonal uncertainty-aware fusion","out-of-distribution detection","road scene understanding","autonomous driving"],"falsifier":"Filter a test set of road scenes to the images where the OUAFS-fused detector outputs no bounding box overlapping a ground-truth anomaly, and measure the cascade's FPR95 and AP on that subset alone. If the cascade is not at least as good as the frozen baseline there, the claim that the framework can boost frozen segmenters is falsified; the paper's own RPL+GD ablation row shows such images exist.","tokens_in":16254,"feed_emoji":"🚗","tokens_out":11905,"duration_ms":114923,"temperature":0.7,"pith_summary":"The paper tries to establish that the two chronic failures of pixel-wise anomaly segmentation — fragmented masks and a one-size-fits-all global score threshold — can be repaired without retraining the pixel-level model. Its recipe is a two-stage cascade: an open-set object detector (GroundingDINO, enhanced by fusing three uncertainty cues under an orthogonality penalty) proposes anomalous regions, and a learned Adaptive Dual-Threshold Network cuts the frozen pixel-wise score map once inside those proposals and once outside them. On the SMIYC and RoadAnomaly benchmarks the cascade lowers false-positive rates by roughly a third to an order of magnitude across baselines (for example RPL's FPR95 from 11.7% to 7.8% on AnomalyTrack and from 0.6% to 0.06% on ObstacleTrack) while raising AP and component-level F1. If correct, the framework is a plug-in upgrade: existing anomaly segmentation models can gain object-level spatial priors by adding the detection stage and the dual thresholds rather than redesigning their scoring. The whole gain, however, is hostage to the detector's proposals — a fact the paper's own ablation illustrates, since bolting on raw GroundingDINO without the two new modules makes every metric worse.","feed_headline":"Region-specific thresholds slash road anomaly false positives","feed_subtitle":"Freezing pixel-wise baselines, a detector-guided dual threshold lifts AP and F1 on three benchmarks.","key_machinery":"The key machinery is the two-stage division of a single thresholding problem into an object-proposal problem and a regional-calibration problem. On the detection side, OUAFS (Orthogonal Uncertainty-Aware Fusion Strategy) serially fuses three uncertainty maps — a semantic segmentation map, a softmax entropy map, and a softmax distance map — with the visual features of GroundingDINO, ordering the fusions by measured contribution and applying an orthogonality loss that penalises redundancy between the fused features. On the segmentation side, ADT-Net (Adaptive Dual-Threshold Network) takes the detector's foreground and background masks and the frozen pixel-wise anomaly score map, normalises scores piecewise against the mean score of each region, and learns two thresholds from architecturally identical predictors, one for foreground and one for background. A cumulative-distribution-function relaxation makes the hard thresholding operation differentiable during training, while a cross-entropy plus threshold-divergence loss keeps the two cuts apart. The mechanism works because it converts an arbitrarily distributed score map into a form where two learned cuts can adapt to each image.","core_discovery":"The paper's central claim, stated on its own terms: anomaly segmentation should be solved coarse-to-fine, with an object-level detector deciding where anomalies are and a pixel-level scorer deciding how much of each detected region is anomalous. The authors claim to be the first to apply an open-set detection model to anomaly segmentation, and their evidence is that the two new modules convert a frozen pixel-wise scorer into a substantially better segmenter. Specifically, OUAFS raises the detector's ability to localize anomalies of all sizes (mAP improves with each added uncertainty cue and with the sequential ordering), and ADT-Net's region-adaptive thresholds, trained with a cumulative-distribution-function relaxation of hard binarization, are what cut FPR95 while preserving recall. The paper reports that across thirteen metrics on three datasets, the framework improved twelve for RPL and eleven for RbA, with the best configuration reaching 98.9% AuROC on RoadAnomaly.","pith_inferences":["Because ADT-Net only re-thresholds regions the detector proposes, the framework's gain should track detector recall: on any image where the OUAFS-fused GroundingDINO outputs no box covering the anomaly, the cascade should collapse toward (or below) the frozen baseline — a testable prediction that the paper's RPL+GD ablation row already gestures at.","The dual-threshold stage is essentially a learned, region-conditional calibration of whatever score map it is given, so the plug-in claim is testable beyond RPL and RbA: the same ADT-Net should improve energy-based or uncertainty-based score maps, provided the proposals cover the anomalies.","Since ADT-Net learns from box proposals and score statistics rather than anomaly annotations, a natural extension is unsupervised region thresholding — training the two predictors on detector pseudo-labels alone would test whether the pixel-wise scorer is needed at all during training.","One internal inconsistency a careful reader will notice: the RPL baseline on AnomalyTrack is reported as 11.7% FPR95 and 83.4% AP in Table 1 but as 7.18% and 88.55% in Table 3, so the exact magnitude of the claimed gains depends on which baseline row is taken as reference; the direction of the gains is consistent, but the size needs a check."],"forward_implications":["Frozen pixel-wise anomaly segmentation models can be upgraded as a plug-in: OoDDINO improves AP and lowers FPR95 for RPL and RbA, and the appendix reports the same transfer to PEBAL and Mask2Anomaly (Mask2Anomaly FPR95 on RoadAnomaly drops from 13.45% to 7.33%).","False-positive control is the main, consistent win: FPR95 falls on every dataset and baseline, and in the small-obstacle and night-time conditions of ObstacleTrack it falls by an order of magnitude (RPL from 0.6% to 0.06%).","Component-level mask quality improves alongside pixel metrics, because the detector's boxes restore object coherence: PPV and F1 rise substantially (RPL F1 from 30.1 to 56.1 on AnomalyTrack and from 56.6 to 86.5 on ObstacleTrack).","The gain decomposes cleanly: ADT-Net alone repairs most of the damage caused by raw detector errors (AnomalyTrack FPR95 from 21.25% down to 6.50%), while OUAFS lifts detector mAP across small, medium, and large anomalies.","The upgrade carries a real compute cost that the paper discloses: the full model uses about 660M parameters and 410.88 GFLOPs versus 168M and 32.1 GFLOPs for RPL, with inference speed dropping from 4.56 to 3.51 FPS."],"supporting_citations":[{"why":"Supplies GroundingDINO, the open-set object detector whose proposals define the foreground and background regions used for dual thresholding.","marker":"[34]"},{"why":"RPL is the frozen pixel-wise baseline whose FPR95 and AP the cascade improves in the main tables.","marker":"[35]"},{"why":"RbA is the second frozen pixel-wise baseline; OoDDINO pushes its RoadAnomaly AuROC to 98.9%.","marker":"[40]"},{"why":"S2M provides the synthetic training data and the argument that object-level spatial priors are needed rather than raw pixel scores.","marker":"[53]"},{"why":"Source of the softmax-distance uncertainty map that OUAFS fuses with entropy and segmentation cues.","marker":"[10]"},{"why":"CF-MAD motivates the multimodal fusion design that OUAFS refines with sequential fusion and an orthogonality constraint.","marker":"[47]"},{"why":"SMIYC (AnomalyTrack and ObstacleTrack) is the main benchmark whose pixel-level and component-level metrics are reported.","marker":"[5]"},{"why":"RoadAnomaly benchmark supplies the 60 real-world images with pixel-level anomaly labels used in the main comparison.","marker":"[33]"},{"why":"PEBAL serves as an additional frozen baseline in the transferability experiments showing the plug-in gains extend beyond RPL and RbA.","marker":"[45]"}],"fun_headline_variants":["Object-level detection sharpens anomaly segmentation","Region-adaptive thresholds reduce anomaly false positives","Uncertainty-guided fusion improves road anomaly localization","Plug-in dual-threshold network boosts anomaly segmentation","Coarse-to-fine framework enhances anomaly segmentation accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework stands or falls on the detector's proposals: if the OUAFS-fused GroundingDINO fails to place a box over an actual anomaly, or places spurious boxes over normal road content, ADT-Net's region-specific thresholds inherit that mistake and the whole cascade can end up worse than the frozen baseline — the paper's own ablation shows raw GroundingDINO integration (without ADT-Net or OUAFS) raising RPL's FPR95 on RoadAnomaly from 17.74% to 28.50%.","fun_headline_variants_meta":{"raw":{"variants":["Object-level detection sharpens anomaly segmentation","Region-adaptive thresholds reduce anomaly false positives","Uncertainty-guided fusion improves road anomaly localization","Plug-in dual-threshold network boosts anomaly segmentation","Coarse-to-fine framework enhances anomaly segmentation accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00029,"raw_usage":{"total_tokens":1729,"prompt_tokens":1010,"completion_tokens":719,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":650}},"tokens_in":626,"tokens_out":719,"duration_ms":8470,"temperature":1.0,"reasoning_tokens":650,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:50:51.205002+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Filter a test set of road scenes to the images where the OUAFS-fused detector outputs no bounding box overlapping a ground-truth anomaly, and measure the cascade's FPR95 and AP on that subset alone. If the cascade is not at least as good as the frozen baseline there, the claim that the framework can boost frozen segmenters is falsified; the paper's own RPL+GD ablation row shows such images exist.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies GroundingDINO, the open-set object detector whose proposals define the foreground and background regions used for dual thresholding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RPL is the frozen pixel-wise baseline whose FPR95 and AP the cascade improves in the main tables."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RbA is the second frozen pixel-wise baseline; OoDDINO pushes its RoadAnomaly AuROC to 98.9%."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"S2M provides the synthetic training data and the argument that object-level spatial priors are needed rather than raw pixel scores."},{"cited_title":"Siegwart, and César Cadena","cited_arxiv_id":null,"evidence_quote":"Source of the softmax-distance uncertainty map that OUAFS fuses with entropy and segmentation cues."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CF-MAD motivates the multimodal fusion design that OUAFS refines with sequential fusion and an orthogonality constraint."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SMIYC (AnomalyTrack and ObstacleTrack) is the main benchmark whose pixel-level and component-level metrics are reported."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RoadAnomaly benchmark supplies the 60 real-world images with pixel-level anomaly labels used in the main comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PEBAL serves as an additional frozen baseline in the transferability experiments showing the plug-in gains extend beyond RPL and RbA."}],"review_version":1}