{"id":"510a0631-d717-49a1-bea8-17150a5582f4","arxiv_id":"2412.10176","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"UN-DETR learns a jointly supervised instance presence score from positional and categorical information, improving unknown object detection over prior objectness methods.","lead":"This paper introduces UN-DETR, a transformer-based detector that learns an 'instance presence score' by combining location and category signals to find objects of unseen classes. The authors report state-of-the-art results on unknown-object detection benchmarks, though key comparison and threshold details remain open.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA margin may be an artifact of unmatched unsupervised pretraining: the paper's own ablation (Table 3, row 8) shows UN-DETR without pretraining is below UnSniffer on U-F1 and U-PRE, so the claimed gains are not attributable to the proposed joint-supervision mechanism.","rationale":"The paper is best read as a system paper: the full pipeline, including unsupervised pretraining, is the proposed method, and SOTA is claimed for the full pipeline. That is legitimate, but the comparison must control for ingredients. The paper's own ablation gives an internal control: without pretraining, the system is not SOTA on the two headline metrics, U-F1 and U-PRE. Since baselines did not receive this pretraining, the comparison is not apples-to-apples. The appendix's retraining attempts are partial and are not integrated into the main tables, so the correction is incomplete. The reader flagged threshold tuning as the weakest assumption; that omission is a real reproducibility issue, but the pretraining confound is more decisive because it is evidenced by the paper's own ablation. I do not see the paper as dishonest; these are addressable experimental-control concerns. The appropriate verdict remains conditional: the authors need to provide a fair pretraining-matched comparison, or explicitly limit the SOTA claim to the full pipeline with their pretraining recipe.","tokens_in":15494,"tokens_out":6756,"duration_ms":64137,"concrete_test":"Retrain UnSniffer (and, if feasible, the other DETR-based baselines) with the same unsupervised pretraining on the VOC training set using Selective Search and SwAV, and re-evaluate on COCO-OOD with identical unknown/known threshold rules; then compare U-F1 and U-PRE to UN-DETR. If UnSniffer with pretraining reaches or exceeds 54.9/54.5, the headline SOTA claim collapses to a pretraining advantage; if UN-DETR still leads by comparable margins, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"UN-DETR's headline claim is that it surpasses the previously leading method (UnSniffer) in UOD, with U-F1 and U-PRE gains of 7.0 and 11.2 points on COCO-OOD. The paper's own ablation in Table 3, row 8, however, shows that removing the unsupervised pretraining component drops U-F1 from 54.9 to 44.5 and U-PRE from 54.5 to 35.0. The same COCO-OOD table lists UnSniffer at U-F1 47.9 and U-PRE 43.3. Thus the full UN-DETR beats UnSniffer only when the pretraining component is included, and no baseline in Tables 1-2 is given this pretraining. The appendix acknowledges the mismatch, reporting retrained OW-DETR and PROB with U-AP gains of only 4.4 and 6.8, and states that UnSniffer with a similar pretraining procedure yielded a final U-AP of 43.9, below its original 45.4; these results are not integrated into the main tables. Consequently, the claimed margin over the previous SOTA is confounded with the pretraining recipe, and the central attribution of the gains to the jointly supervised IPS is not established by the reported experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UN-DETR, a transformer-based unknown object detector built on Deformable DETR. Its central novelty is an Instance Presence Score (IPS) predicted by a dedicated head, trained with joint supervision from positional (GIoU) and categorical (sum of known-class logits) signals, with a one-to-many assignment for additional positive samples. IPS is also used for query selection and post-processing, and the whole detector is unsupervised-pretrained on the training set with region priors. The paper reports state-of-the-art results on COCO-OOD and strong results on COCO-Mixed, and includes ablations of each component.","tokens_in":15832,"tokens_out":4417,"duration_ms":46791,"significance":"If the reported results hold under a fair comparison, this is a meaningful contribution: it is the first transformer-based UOD method, it introduces a learnable objectness score that combines two complementary signals, and it provides a fairly complete ablation study with convergence and visualization analyses. The authors also provide a code repository, which supports reproducibility. However, the central SOTA claim is currently supported by a comparison in which the proposed method uses an unsupervised pretraining recipe that is not applied to the baselines in the main tables, and the paper's own ablation shows the pretraining is largely responsible for the headline U-F1 and U-PRE margins. The contribution is therefore promising but not yet convincingly established.","major_comments":[{"comment":"The claim that UN-DETR surpasses the previous leading method (Table 1: U-F1 54.9 vs 47.9, U-PRE 54.5 vs 43.3 on COCO-OOD) is confounded by the unsupervised pretraining component. Removing pretraining drops U-F1 from 54.9 to 44.5 and U-PRE from 54.5 to 35.0, both below UnSniffer's 47.9 and 43.3. The appendix's retrained-baseline results are not integrated into Tables 1-2, and the reported UnSniffer-with-pretraining U-AP of 43.9 is below its original 45.4. Thus the headline margin is not attributable to the jointly supervised IPS mechanism alone; please integrate the pretrained-baseline comparisons into the main tables, or provide a no-pretraining comparison against all baselines on the full metric set.","section":"Ablation Study, Table 3 row 8"},{"comment":"The known/unknown decision is made by checking whether 'both classification confidence and IPS are above set thresholds', but the threshold values are never reported, and no sensitivity analysis or protocol for choosing them is given. Since U-F1 and U-PRE are computed after this decision rule, the headline results are not independently reproducible from the text. Please report the thresholds, state how they were selected (e.g., fixed in advance or tuned on COCO-OOD), and provide a sensitivity study over these thresholds.","section":"IPS-Guided Post Process, 'Dual-Criteria Unknown Distinguish Protocol'"},{"comment":"The IPS supervision target Po = α·eo_bbox + β·eo_cls is not an independent objectness measurement: eo_cls is the sum of the model's own classification logits for the matched prediction, and eo_bbox is the GIoU between the model's predicted box and the ground-truth box. Using the model's own confidence as a categorical supervision signal is a legitimate training strategy, but the paper should clarify this explicitly and justify why it does not induce confirmation bias in the objectness estimate. A concrete validation would be to compare against an IPS trained with binary foreground/background labels for known objects, or to report the correlation between the learned IPS and GIoU on held-out known objects.","section":"Instance Presence Score Predictor, Eqs. (1)-(4)"}],"minor_comments":[{"comment":"The caption writes 'COCO-ODD'; this should be 'COCO-OOD'.","section":"Table 1 caption"},{"comment":"The text says 'jointly supervised singals from complementary positional and categorical latent space'; 'singals' should be 'signals', and 'latent space' should be plural.","section":"Introduction, contributions"},{"comment":"'Futhermore' should be 'Furthermore'.","section":"Appendix, Unsupervised Pre-training"},{"comment":"The metric is written both as 'GIoU' and 'GIOU'; please unify the notation.","section":"Eqs. (1)-(4)"},{"comment":"The text says 'we introduce an additional IPP' for query selection, but the same acronym IPP is also used for the main objectness head. Please clarify whether these are the same head, separate heads, or shared-parameter heads, and how the two IPP losses interact.","section":"Unbiased Query Selection"},{"comment":"The appendix refers to 'the donut in the second row ... of Figure 3' and 'the moon in the fifth row of Figure 4'; these should reference the qualitative comparison figures (Figures 8 and 9).","section":"Appendix, Qualitative Analysis"}],"recommendation":"major_revision","confidential_remarks":"The central issue is fairness of the comparison. The paper's own ablation row 8 is decisive: without unsupervised pretraining, UN-DETR falls below UnSniffer on U-F1 and U-PRE, the exact metrics used to claim SOTA. The appendix shows that applying similar pretraining to UnSniffer is not straightforward and actually hurts its U-AP, so the problem is not simply 'add pretraining to baselines'; it is that the paper needs to present an evidence chain that attributes the margin to the proposed joint-supervision mechanism rather than to the pretraining recipe. The missing threshold values for the dual-criteria protocol are an additional reproducibility risk. I would support acceptance only after these are addressed with a matched comparison or a clearly stated no-pretraining baseline table."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, short version: this is a serious empirical paper with a genuine new construction, but the headline SOTA margin is larger than the evidence supports because the main comparison is not matched on the unsupervised pretraining component. The paper deserves a referee, but the authors need to redo the comparison before the margin is believable.\n\nWhat's new: they build a Deformable DETR for unknown object detection and add a dedicated Instance Presence Score learned under joint supervision from both box overlap (GIoU) and summed class logits. That head then drives query selection, NMS, and known/unknown separation. The one-to-many assignment for the IPS head, the UQS variant, and the IPS-guided post-processing are each ablated and each shows a gain. That is a coherent, incremental but real contribution. The writing and experiments are honest enough to include the ablation that causes the trouble.\n\nThe soft spot: Table 3 row 8. Removing the unsupervised pretraining (which is DETRreg-style, from Bar et al.) drops U-F1 from 54.9 to 44.5 and U-PRE from 54.5 to 35.0, both below the UnSniffer numbers in Table 1 (47.9 / 43.3). So the full system beats UnSniffer, but the proposed joint-supervision components without pretraining do not. The appendix tries to close the gap by retraining OW-DETR and PROB with the same pretraining, but those only gain a few U-AP points, and UnSniffer with a similar procedure actually drops from 45.4 to 43.9 U-AP. That means the pretraining recipe is not a neutral boost; it helps their architecture and hurts UnSniffer. The claimed margin is therefore confounded with the pretraining choice. The paper's conclusion attributes the gains to jointly supervised IPP, but the ablation doesn't support that attribution as the sole or main driver.\n\nTwo smaller issues: the dual-criteria known/unknown thresholds are never given, and U-PRE/U-F1 are sensitive to them; and there are no error bars, so we don't know if the 7-point margin is stable. The benchmark is also small (504 images for COCO-OOD), which is standard for this subfield but worth remembering.\n\nBottom line: the paper is a useful system paper for anyone working on open-world or unknown object detection. The IPS construction is worth citing. But treat the SOTA claim as conditional until the pretraining-matched comparison is integrated into the main tables and the thresholds are released. I'd send it to review, with a request for those fixes.","headline":"Solid system paper with a real but incremental contribution; the SOTA margin is confounded with unsupervised pretraining and the attribution to the new IPS head doesn't survive the authors' own ablation.","tokens_in":16379,"tokens_out":2513,"would_cite":true,"duration_ms":24819,"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":"Training an objectness score on both box overlap and classification confidence lets a transformer detector outperform prior methods at finding objects of unseen categories.","keywords":["unknown object detection","objectness learning","instance presence score","joint supervision","transformer detector","open-world object detection","unsupervised pretraining","one-to-many assignment"],"falsifier":"The claim would be falsified if re-running the released model on COCO-OOD while sweeping the dual-criteria thresholds reveals that the reported 54.9 U-F1 and 54.5 U-PRE are only obtained at the sweep maximum rather than at a pre-fixed threshold setting.","tokens_in":15296,"feed_emoji":"🎯","tokens_out":10293,"duration_ms":91282,"temperature":0.7,"pith_summary":"The paper argues that prior unknown-object detectors learn objectness from only one source of supervision, either box overlap or classification confidence, and that this split explains their poor recall or precision on unseen objects. It proposes UN-DETR, a transformer-based detector that trains a dedicated Instance Presence Score (IPS) under joint supervision from both spaces: a generalized-IoU term from the regression head and the summed foreground logits from the classification head. This score is then reused in query selection, in an IPS-guided NMS, and in a dual-criteria rule that separates known from unknown objects. On the COCO-OOD benchmark the method raises unknown F1 and precision by 7.0 and 11.2 points over the previous best, while roughly preserving known-class mAP, which is what the authors offer as evidence that the joint-supervision strategy is the cause.","feed_headline":"Fused objectness score beats prior methods on unseen-object detection","feed_subtitle":"Training object presence on box overlap plus classification confidence lifts unknown-object F1 and precision to new highs","key_machinery":"Instance Presence Score (IPS), a learnable scalar output by a dedicated one-layer IPS Predictor head, is the load-bearing object. Its training target is $P_o(e_i)=\\alpha\\,\\mathrm{GIoU}(b_i,\\hat{b}_{\\sigma_{pos}})+\\beta\\,\\hat{P}^f_{\\sigma_{pos}}$, with $\\alpha=0.6$, $\\beta=0.4$; the first term is the generalized IoU between the regression head's box and the matched ground-truth box, and the second is the sum of the known-class logits, a foreground probability. Samples whose $\\mathrm{GIoU}$ falls at or below $\\tau=0.6$ are pushed toward a constant $C=0.5$ instead, giving the score a clean separation between foreground-like and background-like boxes. IPS is reused in three places: a one-to-many assignment adds a second set of matched queries as extra positives for IPS training only, Unbiased Query Selection replaces the two-stage DETR query-scoring classification head with an additional IPP head, and an IPS-guided post-process ranks boxes with Distance-IoU NMS before applying a dual-criteria rule (classification confidence and IPS both above thresholds $\\Rightarrow$ known; low classification but high IPS $\\Rightarrow$ unknown).","core_discovery":"The central discovery is that objectness for unknown objects is better learned as a recoupling of two complementary latent spaces rather than from either alone. The paper shows that using only categorical information yields clear foreground/background separation but poor instance separation, while using only positional information yields the opposite. IPS is defined as the probability of an object's presence and is trained to match a convex combination of GIoU(box, matched GT) and the sum of classification logits over known categories, with a low constant target for low-overlap samples. The same IPS is then used to select decoder queries and to guide post-processing, including a known/unknown decision rule based on both classification confidence and IPS. Under the UOD evaluation protocol, this design claims the highest reported unknown-object F1 and precision on COCO-OOD and COCO-Mixed while keeping known-object mAP roughly unchanged.","pith_inferences":["A natural stress test the paper does not run is a full precision-recall sweep of the dual-criteria thresholds; if the reported COCO-OOD numbers sit at the peak of that sweep, the benchmark gains are partly tuning, not pure generalization.","The same joint-supervision recipe could be dropped into open-vocabulary detectors as an auxiliary objectness head, so that objects outside the prompt vocabulary still get localized and marked unknown rather than missed.","The COCO-Mixed results are more modest than the COCO-OOD ones: the method leads in U-AP and U-PRE but trails on U-F1, so a careful reader should weigh how much of the 'surpassing previous methods' claim is specific to the all-unknown test set.","Replacing the fixed convex weights $\\alpha=0.6$, $\\beta=0.4$ with a learned gating of the two signals is a testable variant that might reduce sensitivity to the choice of known categories."],"forward_implications":["If the reported gains hold, augmenting any detector that has separate classification and regression heads with a jointly supervised IPS head should improve unknown-object recall and precision without retraining the backbone from scratch.","Because IPS is class-agnostic, using it in query selection should make the decoder's initial queries less biased toward known categories, which is a direct design lesson for open-world detectors.","The ablation showing U-PRE drops from 54.5 to 35.0 without unsupervised pretraining implies that an objectness prior acquired from region proposals is a major share of the method's precision advantage.","The dual-criteria rule (known if both classification confidence and IPS are high, unknown if only IPS is high) provides a template for separating known from unknown instances using two complementary confidence signals."],"supporting_citations":[{"why":"Provides the Deformable DETR base architecture and the two-stage query selection that Unbiased Query Selection modifies.","marker":"(Zhu et al. 2020)"},{"why":"Defines the UOD benchmark and the localization-only GOC objectness baseline that the paper aims to surpass on COCO-OOD.","marker":"(Liang et al. 2023)"},{"why":"Represents the classification-only objectness approach (PROB) that IPS is designed to complement.","marker":"(Zohar, Wang, and Yeung 2023)"},{"why":"Establishes the end-to-end transformer detection paradigm on which UN-DETR is built.","marker":"(Carion et al. 2020)"},{"why":"Supplies the unsupervised pretraining with region priors that gives the model its objectness prior.","marker":"(Bar et al. 2022)"},{"why":"Selective Search is the region proposal generator used in the unsupervised pretraining.","marker":"(Uijlings et al. 2013)"},{"why":"SwAV is the self-supervised encoder used to align object embeddings during pretraining.","marker":"(Caron et al. 2020)"},{"why":"Distance-IoU is the overlap measure used to rank boxes in the IPS-guided NMS.","marker":"(Zheng et al. 2020)"}],"fun_headline_variants":["Joint supervision of objectness tops unseen-object detection","UN-DETR fuses box and class cues for unknown-object F1","Objectness from two worlds: UN-DETR sets new UOD records","IPS score unifies positional and categorical cues for UOD","Joint supervision for objectness: UN-DETR's key to unknown objects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central premise is that an objectness score trained as a weighted blend of box overlap (GIoU) and the summed known-class logits transfers to unseen categories, and that the known/unknown thresholds in the final protocol are fixed rather than tuned to each test set.","fun_headline_variants_meta":{"raw":{"variants":["Joint supervision of objectness tops unseen-object detection","UN-DETR fuses box and class cues for unknown-object F1","Objectness from two worlds: UN-DETR sets new UOD records","IPS score unifies positional and categorical cues for UOD","Joint supervision for objectness: UN-DETR's key to unknown objects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000599,"raw_usage":{"total_tokens":2815,"prompt_tokens":973,"completion_tokens":1842,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":1751}},"tokens_in":589,"tokens_out":1842,"duration_ms":12962,"temperature":1.0,"reasoning_tokens":1751,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:14:56.242048+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The claim would be falsified if re-running the released model on COCO-OOD while sweeping the dual-criteria thresholds reveals that the reported 54.9 U-F1 and 54.5 U-PRE are only obtained at the sweep maximum rather than at a pre-fixed threshold setting.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the UOD benchmark and the localization-only GOC objectness baseline that the paper aims to surpass on COCO-OOD."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Represents the classification-only objectness approach (PROB) that IPS is designed to complement."},{"cited_title":"J.; Herzig, R.; Chechik, G.; Rohrbach, A.; Darrell, T.; and Globerson, A","cited_arxiv_id":null,"evidence_quote":"Supplies the unsupervised pretraining with region priors that gives the model its objectness prior."},{"cited_title":"R.; Van De Sande, K","cited_arxiv_id":null,"evidence_quote":"Selective Search is the region proposal generator used in the unsupervised pretraining."}],"review_version":1}