{"id":"1c7785c2-a3ff-48a9-b14d-1ab5fad4cc0d","arxiv_id":"2412.03986","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"UNCOVER extends a real-time YOLO detector with an occupancy score, an extra OOD class, and a depth-based filter to catch objects outside the standard traffic classes.","lead":"This paper adds an 'unknown object' detector to a fast self-driving car vision model, so the model can flag things like animals, trailers, or road debris that are not in the usual eight traffic classes. The added parts are small enough that the model still runs in real time, and a depth-based filter removes false alarms from shadows or road markings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'unknown' test categories overlap with the COCO/LVIS auxiliary training data, so the reported recall gains do not isolate whether the occupancy score transfers to genuinely novel objects.","rationale":"The reader's CONDITIONAL verdict is appropriate, and their weakest assumption is the same one I would flag. The paper's central contribution is not merely a detector that recognizes a broad set of classes; it is the claim that the occupancy score is a generic objectness measure generalizing to unknown classes. That claim requires the evaluation to use unknown objects not present in the COCO/LVIS auxiliary data. The current protocol does not satisfy this: traffic lights and signs, used as unknowns on BDD100k, are in the auxiliary data; many Cityscapes 'dynamic' objects and the Anomaly giraffe also appear in COCO/LVIS. This blurs the empirical separation between 'learned category-specific OOD class' and 'learned generic occupancy.' The concern is load-bearing because the OOD class alone could produce high recall for seen categories, while the occupancy head's added value specifically depends on detecting categories not seen. I find no internal inconsistency or mathematical error in the method; the architecture is plausible, the ablations in Table 6 are internally consistent, and the runtime comparison supports the real-time claim. The issue is addressable by a per-category overlap analysis or a held-out novel-object benchmark. If the non-overlap split preserves the occupancy advantage, the central claim is supported; if not, the paper overstates its evidence, though the system may still be practically useful for the AD categories it evaluates. The DFR averaging concern noted by the reader is secondary; the transfer question is the more fundamental one. I therefore keep the CONDITIONAL verdict rather than moving to ACCEPT or REJECT, and the concrete test above would settle the uncertainty.","tokens_in":18650,"tokens_out":7161,"duration_ms":68213,"concrete_test":"Partition every unknown-object ground-truth instance in BDD100k, Cityscapes, FS L&F, and Anomaly by whether its semantic label (e.g., 'traffic light', 'dog', 'giraffe') occurs in the COCO/LVIS category lists used for auxiliary training. Recompute R@100, and the Table 1 occupancy-versus-Obj/IoU comparison, separately for overlap and non-overlap subsets. If the non-overlap subset retains the reported occupancy advantage over YOLO-World and over Obj/IoU baselines, the transfer claim is supported; if the advantage disappears or reverses, the headline recall is largely explained by exposure to those classes rather than by generic objectness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the occupancy head learns a generic, class-agnostic objectness that transfers to objects outside all training data (Sections 3.2.2 and 4.1). The evaluation does not isolate this transfer. In Appendix B.1, the authors explicitly remove 'traffic sign' and 'traffic light' from the known classes during training and use them as unknowns on BDD100k; however, traffic light is a COCO class and traffic sign/stop sign appear in LVIS, so detections of these categories can be produced by the extra OOD class trained on exactly those data. Similarly, many Cityscapes 'dynamic' objects (animals, bags, buggies, trash bins) and the Anomaly benchmark's giraffe are present in COCO/LVIS. Thus Table 4's recall comparison (e.g., up to 25% over YOLO-World on FS) may confound category recognition from the OOD class with the occupancy score's generic objectness. What would have to be true for the central claim is that occupancy produces high scores for boxes around objects not present in AD, COCO, or LVIS; this remains untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UNCOVER, an extension of a real-time one-stage object detector (YOLOX) for detecting unknown-class objects in autonomous driving scenes. The two main contributions are (i) an extra OOD class trained on auxiliary data from COCO and LVIS using a modified Mosaic/Mixup augmentation, together with a regression-head \"occupancy\" score that predicts the fraction of a predicted bounding box covered by any ground-truth box, and (ii) a post-hoc depth-based filtering step (DFR) that uses morphological operations and Sobel gradients on depth maps to reject detections whose boxes lack geometric consistency. The method is evaluated on Cityscapes, BDD100k, Fishyscapes Lost & Found, and the SMIYC Anomaly/Obstacle tracks, reporting recall improvements over YOLO-World (up to 25%), competitive results against anomaly segmentation baselines at much higher throughput, and an average 18.4% FPR reduction with 4.1% recall improvement when applying DFR.","tokens_in":18864,"tokens_out":5343,"duration_ms":50061,"significance":"If the central claim holds, UNCOVER is a practically valuable contribution: it adds unknown-object awareness to a real-time detector with only a small complexity increase, introduces an interpretable depth-based false-positive filter, and provides a convenient evaluation protocol by converting anomaly-segmentation masks to bounding boxes. The paper is also transparent in several ways: Table 6 ablates both the OOD class and the occupancy head, Appendix A.3 and C.7 give threshold-sweep and reproducibility tables, and the supplementary material documents dataset preprocessing in detail. However, the evaluation has a load-bearing blind spot: the benchmark \"unknown\" categories substantially overlap with the COCO/LVIS data used to train the OOD class, so the reported recall gains do not cleanly demonstrate that the occupancy score transfers to genuinely novel objects.","major_comments":[{"comment":"The central claim that the occupancy head learns a generic, class-agnostic objectness that transfers to objects outside all training data is not cleanly tested. Appendix B.1 explicitly removes traffic sign and traffic light from the known classes during training and treats them as unknowns on BDD100k, but both classes appear in COCO and LVIS, which are used as auxiliary OOD training data. Likewise, many other benchmark unknowns (animals, bags, buggies, trash bins, the Anomaly giraffe) are present in COCO/LVIS. Consequently, the Table 4 recall improvements (e.g., up to 25% over YOLO-World) could be driven by recognition of seen OOD categories rather than by the occupancy score's generic objectness. Please re-evaluate with a hold-out split that excludes all test categories (and their near variants) from COCO/LVIS, or at least report recall separately for categories that are present versus absent in the auxiliary training data.","section":"§4.1, Appendix B.1, Table 4"},{"comment":"The recall metric R@100 is computed without exhaustive unknown-object annotations, which makes the headline recall numbers difficult to interpret. The paper itself states in the Figure 9 caption that \"Average precision would penalize the right object detections due to lack of exhaustive object annotations,\" confirming that many true unknowns are unlabeled. Without exhaustive annotations, unlabeled objects can be counted as false positives, and the denominator for recall (the number of ground-truth unknowns) is underestimated, so the reported R@100 is at best an approximation rather than a true recall. Please either construct an evaluation set with exhaustive per-frame unknown annotations (for example, using only categories with full instance-level annotation) or report precision/F1 alongside recall and quantify the estimated proportion of unannotated unknowns.","section":"§4, Evaluation Metrics; Figure 9 caption"},{"comment":"The abstract claims that DFR \"reduces false positive rate by 18.4% while boosting recall by 4.1%,\" but this is an average over a heterogeneous set of methods and datasets, and several individual cells show the opposite behavior. For example, UNCOVER (BDD) on Cityscapes has FPR@100 increase from 2.2 to 2.5 (+13.6%), and the same row on Fishyscapes L&F shows FPR rising from 0.5 to 2.8 despite the parenthetical in the table. The average improvement is thus not a consistent effect. Please report per-configuration changes, use a more robust summary statistic, and reconcile the negative cells with the abstract's claim.","section":"Table 3, Abstract"},{"comment":"All results are reported from a single run with no error bars or multiple seeds, even though several key comparisons are within a few percentage points (e.g., Table 1: Obj. score 13.21 vs. IoU 11.70 vs. Occ 15.71 for Cityscapes R@100). In addition, the thresholds µocc and µ are grid-searched on the evaluation benchmarks (Appendix A.2, A.3). Without variance estimates or a clearer separation between validation and test thresholds, the robustness of the reported improvements is uncertain. Please provide mean±std over at least three seeds for the central comparisons, or a per-dataset threshold-sensitivity analysis with thresholds selected on held-out data only.","section":"§4, Implementation Details"}],"minor_comments":[{"comment":"There is a typo: \"reduce false the positive rate\" should be \"reduce the false positive rate.\"","section":"Abstract"},{"comment":"The notation \"$S^n_{i=0} bgti$\" is nonstandard; it should be \"$\\bigcup_{i=0}^n bgti$\" for the union of ground-truth boxes.","section":"§3.2.2, Eq. (1)"},{"comment":"The bounding-box slicing \"bbox ← C[x1 : y1, x2 : y2]\" appears to use x and y inconsistently with the argument order (x1, x2, y1, y2); please clarify the coordinate conventions.","section":"Algorithm 1"},{"comment":"The parenthetical relative changes are confusing, especially for UNCOVER (BDD) on FS L&F where the FPR increases from 0.5 to 2.8 but the table lists \"-44.0%\"; please correct the sign or explain what is being computed.","section":"Table 3"},{"comment":"References [2] and [3] are duplicates of the same YOLOv4 citation and should be merged.","section":"References"},{"comment":"Typo: \"auxilary\" should be \"auxiliary.\"","section":"Appendix B.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-scoped for a computer-vision venue and the authors have made a reasonable effort in ablation and reproducibility. The main concern is not the method's internal consistency but the benchmark design: because the auxiliary OOD training data (COCO/LVIS) contains many of the evaluation categories, the reported recall gains do not isolate the claimed transfer to genuinely novel objects. This is a structural issue with the evaluation protocol, not a minor wording problem. I would ask the authors to rerun at least a subset of the experiments with a strict OOD split (e.g., withholding a set of COCO/LVIS classes from training and using those as the only evaluation unknowns), and to report recall exhaustion statistics. The DFR average claim also needs to be reconciled with per-cell degradations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the occupancy prediction in Eq. (1) is a genuinely new objectness signal compared to the IoU and Obj scores used in prior work, and the depth-based false-positive filter is a practical post-hoc addition. But the evaluation doesn't isolate the core claim that occupancy transfers to truly novel objects. Many of the 'unknown' test classes (traffic signs, lights, animals, lost cargo) are present in the COCO/LVIS data used to train the OOD class, so the recall gains could largely be category recognition, not generic objectness. That leak is the main soft spot.\n\nWhat's good: the method is clearly described, the ablation in Table 6 shows both the OOD class and occupancy contribute, and it runs at 26 FPS while beating YOLO-World on recall without hurting known-class mAP. The depth filter is interpretable and they report per-dataset numbers rather than only a cherry-picked average. They also acknowledge the recall metric's limitation regarding annotation exhaustiveness.\n\nSoft spots in proportion: the benchmark leak is load-bearing. The paper argues occupancy is class-agnostic but never tests it on objects provably absent from all training data. There are no error bars or multiple seeds, which makes the small recall differences in Table 1 hard to interpret. FPR@100 is valid only inside an exhaustively annotated ROI, so it is not a global false-positive measure. The headline DFR average of -18.4% FPR hides a couple of regressions (e.g., UNCOVER on Cityscapes FPR actually increases), though the table is honest about it. No code released.\n\nThis paper deserves a serious referee. The idea is worth engaging with and the engineering is solid, but the evaluation needs a clean separation between known, OOD-trained, and truly novel categories, plus seeds and ideally code. I would send it to review with a request for major revision.","headline":"Solid real-time unknown-object detector with a new occupancy signal and useful depth filter, but the headline recall gains don't isolate true novelty because test unknowns overlap with the training auxiliaries.","tokens_in":19432,"tokens_out":3233,"would_cite":false,"duration_ms":30540,"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 real-time detector can be taught a class-agnostic sense of objectness from occupancy and flag unknown road hazards with it.","keywords":["unknown object detection","out-of-distribution detection","autonomous driving","real-time object detection","objectness","occupancy prediction","depth-based filtering","open-world object detection"],"falsifier":"Train UNCOVER on Cityscapes plus COCO/LVIS, then test on a set of driving images into which objects from classes absent from all four datasets (for example, furniture or office items) have been composited with ground-truth boxes. Measure R@100 with the paper's thresholds; if recall collapses to near zero on these truly novel classes while staying high on COCO/LVIS-style unknowns, the claimed transfer of occupancy to novel objects is not supported.","tokens_in":18387,"feed_emoji":"🚗","tokens_out":8470,"duration_ms":68253,"temperature":0.7,"pith_summary":"UNCOVER asks whether a real-time object detector can flag objects it was never trained on, and answers yes by replacing the usual localization-quality objectness score with an occupancy score: the fraction of the predicted box that overlaps any ground-truth box at all. Trained with an extra OOD class and with Mosaic+ and Mixup augmentation over COCO and LVIS images, the model learns a class-agnostic 'something is here' signal that transfers to unknowns. On Cityscapes, BDD100k, Fishyscapes, and SegmentMeIfYouCan it reports unknown-object recall that beats the real-time open-vocabulary detector YOLO-World by up to 25 percentage points while staying at about 26 FPS on a V100. A post-hoc depth-based filter then reduces false positives, mainly at close range, by checking whether a detection has the depth-change signature of a physical object.","feed_headline":"New occupancy score lets real-time detectors spot unknown road objects","feed_subtitle":"UNCOVER adds one regression head and a depth filter, lifting unknown-object recall up to 25 points over YOLO-World.","key_machinery":"The central object is the occupancy prediction defined by the training target in Eq. (1): for a predicted box $b_{\\text{pred}}$ and the union of all ground-truth boxes $\\cup_i b_{gt_i}$, the target is $t_{occ} = |b_{\\text{pred}} \\cap (\\cup_i b_{gt_i})|\\,/\\,|b_{\\text{pred}}|$, trained with binary cross-entropy. Unlike the YOLOX objectness score or the IoU score, this target does not require the predicted box to match a single ground-truth box of the correct class; any overlap with any object counts, which makes the score class-agnostic and tolerant of poor localization. At inference, the occupancy output is used as a second filter: detections with classification score below threshold but occupancy above $\\mu_{occ}$ are kept as OOD objects. The other load-bearing pieces are Mosaic+, which takes two of the four mosaic tiles from COCO/LVIS so the model sees objects outside the AD label space (with Mixup blending the composed image with an AD image to reduce the domain gap), and the depth-based post-filter (Algorithm 1), which applies morphological closing followed by a Sobel gradient in the y-direction to the depth map and keeps a detection when the fraction $c$ of low-depth-change pixels in its box is at least $\\mu$.","core_discovery":"The paper's central claim is that predicting occupancy—the ratio of the predicted box's area covered by the union of all ground-truth boxes—teaches a detector a generic objectness that survives class boundaries. The standard YOLOX objectness score and the IoU score are supervised by box matching and therefore reward precise localization of known classes; the occupancy target instead rewards boxes that simply contain object pixels, known or not. With this target, an extra OOD class, and COCO/LVIS data mixed in via Mosaic+ and Mixup, UNCOVER reports unknown-object recall of 15.71% on Cityscapes, 39.42% on BDD100k, 58.56% on Fishyscapes Lost & Found, and 93.75%/77.78% on the Anomaly/Obstacle tracks, compared with 12.23%, 29.24%, 33.70%, 100%/62.22% for YOLO-World. It also preserves known-class mAP, and the depth-change post-filter improves FPR@100 by 18.4% and R@100 by 4.1% on average.","pith_inferences":["A limitation the paper itself notes: many benchmark 'unknowns' (traffic signs, animals, lost cargo) also occur in the COCO/LVIS training boxes, so the reported recall may overstate transfer to genuinely novel categories; a test set whose classes are provably absent from all training data would settle this.","The same occupancy target could be applied to mask-based heads, replacing box overlap with mask overlap to give an instance-aware objectness signal for unknown-instance segmentation without a heavy decoder.","Because the depth filter only removes detections, its recall gains are partly a reallocation of the top-100 budget; a system with a larger budget or different ranking would likely see different FPR/R trade-offs, so the filter should be re-tuned at the system level."],"forward_implications":["Autonomous-driving detectors can gain unknown-object awareness with only one extra classification logit and one occupancy regression output, keeping real-time throughput (about 26 FPS on a V100 in the paper's measurements).","The occupancy score works as a class-agnostic second filter, so detections with low classification confidence but high occupancy are kept as OOD rather than discarded.","The depth-based post-filter reduces near-range false positives from shadows and flat-surface markings, and it can be attached to any existing detector that has access to a depth map.","Anomaly segmentation benchmarks can be converted to box-level evaluation, giving object-detection researchers a way to measure unknown recall on Fishyscapes and SegmentMeIfYouCan."],"supporting_citations":[{"why":"YOLOX is the one-stage anchor-free base detector whose objectness score and decoupled heads UNCOVER modifies.","marker":"[14]"},{"why":"MS COCO supplies the auxiliary object annotations used to train the extra OOD class and to diversify Mosaic+ inputs.","marker":"[26]"},{"why":"LVIS provides the larger vocabulary of annotated objects that the paper credits for learning a generic sense of objectness.","marker":"[15]"},{"why":"Introduced IoU-based objectness for open-world proposals, the measure UNCOVER compares against and replaces with occupancy.","marker":"[21]"},{"why":"Prior FCOS-based real-time open-world detector that uses localization-quality scores; UNCOVER extends this line with occupancy prediction.","marker":"[23]"},{"why":"YOLO-World is the real-time open-vocabulary baseline whose unknown recall UNCOVER claims to beat by up to 25 percentage points.","marker":"[6]"},{"why":"Cityscapes is the primary AD training set and the source of the reserved unknown classes (dynamic, trailer, caravan) used for evaluation.","marker":"[7]"},{"why":"BDD100k provides the second AD training/evaluation set, with traffic signs and traffic lights held out as unknowns.","marker":"[38]"},{"why":"Sobel operator is the gradient filter used to compute the depth-change map in Algorithm 1.","marker":"[33]"},{"why":"Morphological closing (dilation and erosion) pre-processes the depth map before the Sobel step in the depth filter.","marker":"[11]"}],"fun_headline_variants":["Occupancy target teaches detectors to see unknown objects","UNCOVER: real-time unknown-object detection via occupancy","Generic objectness: occupancy prediction in UNCOVER","Spot unseen road objects in real time with occupancy scoring","Unknown objects? Occupancy scoring for real-time AD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on occupancy learned from the union of AD and COCO/LVIS boxes transferring to objects unlike anything in those sets, yet the evaluation's 'unknown' classes, such as traffic signs, animals, and lost cargo, largely appear in the very COCO/LVIS boxes used for training.","fun_headline_variants_meta":{"raw":{"variants":["Occupancy target teaches detectors to see unknown objects","UNCOVER: real-time unknown-object detection via occupancy","Generic objectness: occupancy prediction in UNCOVER","Spot unseen road objects in real time with occupancy scoring","Unknown objects? Occupancy scoring for real-time AD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001003,"raw_usage":{"total_tokens":4263,"prompt_tokens":984,"completion_tokens":3279,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":3203}},"tokens_in":600,"tokens_out":3279,"duration_ms":23304,"temperature":1.0,"reasoning_tokens":3203,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:52:48.119354+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train UNCOVER on Cityscapes plus COCO/LVIS, then test on a set of driving images into which objects from classes absent from all four datasets (for example, furniture or office items) have been composited with ground-truth boxes. Measure R@100 with the paper's thresholds; if recall collapses to near zero on these truly novel classes while staying high on COCO/LVIS-style unknowns, the claimed transfer of occupancy to novel objects is not supported.","supporting_citations":[{"cited_title":"European conference on computer vision pp","cited_arxiv_id":null,"evidence_quote":"MS COCO supplies the auxiliary object annotations used to train the extra OOD class and to diversify Mosaic+ inputs."},{"cited_title":"IEEE Robotics and Automation Letters7, 5453–5460 (2021)","cited_arxiv_id":null,"evidence_quote":"Introduced IoU-based objectness for open-world proposals, the measure UNCOVER compares against and replaces with occupancy."},{"cited_title":"Extending One-Stage Detection with Open-World Proposals","cited_arxiv_id":"2201.02302","evidence_quote":"Prior FCOS-based real-time open-world detector that uses localization-quality scores; UNCOVER extends this line with occupancy prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cityscapes is the primary AD training set and the source of the reserved unknown classes (dynamic, trailer, caravan) used for evaluation."},{"cited_title":"a talk at the Stanford Artificial Project in pp","cited_arxiv_id":null,"evidence_quote":"Sobel operator is the gradient filter used to compute the depth-change map in Algorithm 1."},{"cited_title":"Eng- land: John Wiley & Sons Ltd pp","cited_arxiv_id":null,"evidence_quote":"Morphological closing (dilation and erosion) pre-processes the depth map before the Sobel step in the depth filter."}],"review_version":1}