{"id":"1def712d-eb1f-4f31-81b0-de47ab0e7f10","arxiv_id":"2411.13001","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A detector trained with feature contrastive and uncertainty classification losses improves open-set semi-supervised object detection, labeling unseen objects as unknown.","lead":"The paper adds two training losses to a semi-supervised object detector so it can recognize objects from categories it never saw labeled, marking them as unknown instead of forcing them into known classes. The reported gains are a few mAP points plus a new capability to flag out-of-distribution objects.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OOD signal hinges on unverified top-K background mining; no evidence that OOD objects are high-background proposals, and K is undefined.","rationale":"The reader's weakest assumption identifies the same load-bearing condition: the method's OOD detection capability rests entirely on the teacher surfacing OOD objects as high-background proposals. I agree with that reading and sharpen it by noting that K is never specified and that the paper's own qualitative example shows a failure mode where an OOD object is confidently classified as ID by the baseline teacher. Without a quantitative recall analysis of the mined background proposals, the reported AP_u could be an artifact of the mining procedure rather than evidence of open-set generalization. I also considered two secondary concerns: the missing DCO OSSOD baseline on the COCO-Open settings and the incomplete loss specification in Eq. 3. Those are real but less central than the mining assumption, because even a complete equation set and a fair baseline would not rescue the method if the OOD training signal is absent. The appropriate verdict remains CONDITIONAL: the central claim is plausible but unverified until the mining step is shown to actually surface OOD objects. Since my assessment does not change the reader's verdict, I leave it UNCHANGED.","tokens_in":11915,"tokens_out":8330,"duration_ms":87077,"concrete_test":"Re-run the CFL pipeline on COCO-Open-CLS with 20 ID classes, logging the teacher's background-class proposals at the start of Stage 2. For K in {100, 300, 1000, 3000}, match mined proposals to ground-truth OOD boxes using IoU > 0.5 and report recall@K and precision; also report the fraction of ground-truth OOD boxes whose teacher softmax argmax is an ID class. If OOD recall@K is below about 50% or the ID-argmax fraction is high, the top-K mining assumption fails and the AP_u increase needs a different explanation. Releasing code with the exact K and mining statistics would also settle the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central OOD-learning mechanism is the top-K background mining in Sec. 4.4: \"we mine potential OOD instances by selecting the top K candidates from the background class probabilities\" (Eq. 5). Everything the model learns about 'unknown' comes from these proposals, because the semi-supervised branch of Eq. 7 contains only -sum w_u log(p_u). If OOD objects are instead confidently assigned to ID classes by the teacher, they never enter this mined set, and L_uc teaches the model to relabel true background as 'unknown' rather than to detect OOD. The paper gives no measurement of how often OOD ground-truth boxes appear among the teacher's high-background proposals, and the paper's own Fig. 4 shows the baseline teacher labeling an OOD 'horse' as 'cow' with high confidence, so the assumption is not self-evidently satisfied. The parameter K is also never specified, so the size and precision of the mined OOD set are unreproducible. If recall@K of true OOD boxes is low, the reported AP_u values (5.56, 7.11, 9.82) would not reflect genuine open-set generalization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CFL-Detector, an open-set semi-supervised object detection method built on a Faster R-CNN / Unbiased Teacher pipeline. It adds a feature contrastive loss (L_fc) that clusters embeddings of the same class and separates different classes, and an uncertainty classification loss (L_uc) that trains the detector to output a unified 'unknown' class for out-of-distribution objects. OOD supervision is obtained by mining the top-K proposals with the highest background-class probabilities in the teacher's predictions. Experiments on COCO-Open-CLS, COCO-Open-SUP, and VOC-COCO report consistent improvements in ID mAP over UT and STAC and substantially higher OOD AP_u, with ablations attributing the gains to the two proposed losses and showing transfer to STAC. The central idea is plausible and the empirical tables are internally consistent, but several load-bearing details of the training objective and the OOD-mining procedure are underspecified.","tokens_in":12223,"tokens_out":8361,"duration_ms":79519,"significance":"If the reported results hold, the method is a simple and practical contribution: it converts a closed-set SSOD detector into an open-set detector by mining background proposals for OOD training and adding a feature-space contrastive term, without requiring an offline OOD detector. The three benchmark settings and the transfer ablation to STAC give the central claim reasonable empirical support. The main caveat is that all OOD supervision is manufactured from top-K background mining, and the paper provides no precision/recall evaluation of that mining step; combined with missing details in the loss definitions, this makes the exact training recipe hard to reconstruct and the AP_u numbers difficult to interpret. The method is not a derivation, but the empirical claims are clear and the issues are addressable in revision.","major_comments":[{"comment":"The supervised loss in Eq. (3) is written as L_sup = L_rpn_cls + L_rpn_reg + L_roi_reg + alpha_t L_fc + beta L_uc, so it contains no ROI-head classification loss L_roi_cls. The sentence immediately below the equation says 'L_cls represents the classification loss,' but no such term appears. The uncertainty loss in Eq. (7) does include a weighted ID term w_k log(p_k), so it may be intended to replace the standard classifier loss, but then it also includes an OOD term w_u log(p_u) on labeled data where there are no OOD labels. Please specify the exact supervised objective, including how w_u is obtained on labeled data, and confirm whether the standard L_roi_cls is used or deliberately omitted. This is load-bearing because the ID mAP improvements over UT could be due to a different training recipe rather than the proposed losses.","section":"§4.2, Eq. (3)"},{"comment":"In the semi-supervised branch, L_uc is reduced to -sum w_u log(p_u), and L_unsup in Eq. (3) contains only RPN classification, L_fc, and L_uc. Thus, on unlabeled data the student receives no ROI-head ID classification loss and no box regression loss. This is a major deviation from the Unbiased Teacher baseline and from the standard pseudo-label pipeline. The paper should either add the usual supervised losses on pseudo-labeled proposals or provide an ablation showing that the proposed losses alone are sufficient; otherwise the reported ID mAP gains and the bounding-box quality in Fig. 4 are unexplained. If regression is intentionally disabled because pseudo-boxes are noisy, state that explicitly and quantify its effect.","section":"§4.4, Eq. (7)"},{"comment":"The only source of OOD supervision in the semi-supervised stage is the set of top-K proposals mined from the background class probabilities, but the value of K is never given in Sec. 5.1 or in any table, and no analysis reports how many of these K proposals are true OOD objects rather than background. If OOD objects are confidently assigned to ID classes by the teacher, they will not enter the mined set, and L_uc will teach the model to call true background 'unknown.' Please report K and provide OOD recall/precision of the mining step (e.g., what fraction of ground-truth OOD boxes appear among the top-K background proposals) on COCO-Open-CLS. Without this, the AP_u numbers in Tables 1 and 2 cannot be attributed to genuine open-set generalization.","section":"§4.4, Eq. (5), Sec. 5.1"},{"comment":"The feature contrastive loss L_fc requires class labels c_i for every proposal and uses S_IoU and S_cos thresholds to update the memory pool. In the unlabeled stage there are no ground-truth boxes, so it is unclear how c_i, the IoU scores, and the 'unknown' features are obtained for the memory pool. Please specify the pseudo-label generation and filtering procedure for the unsupervised contrastive branch (teacher predictions? RPN proposals with class scores? class centers from the memory pool?), and explain how OOD features are selected for storage. This is necessary to reproduce the method and to interpret the ablation in Table 4a.","section":"§4.3, Eq. (4)"},{"comment":"The normalization in Eq. (5) depends on c: for ID classes the denominator includes all classes, while for the OOD class it includes only background and OOD. As a result, p_u and p_k are not probabilities from the same softmax distribution, and the weighted cross-entropy terms in Eq. (7) do not correspond to a single classifier output. Please clarify the intended probabilistic interpretation or compare empirically with a single softmax over {background, ID classes, unknown}. If the asymmetric normalization is intentional, a formal justification is needed because this is the core OOD classification loss.","section":"§4.4, Eq. (5)"}],"minor_comments":[{"comment":"'Task Defination' should be 'Task Definition'.","section":"§3.1"},{"comment":"'scherer et al.' should be 'Scheirer et al.'.","section":"§2.1"},{"comment":"The row labels w_kL_ce and w_uL_ce do not map cleanly to Eq. (7); state which terms from Eq. (7) are disabled in each row and give the value of alpha used in w_u.","section":"Table 5a"},{"comment":"The figure is very small and the arrows/labels ('transfer', 'Select', 'Logits', 'uncertainty estimate') are hard to follow; please enlarge it and use terminology consistent with Secs. 4.3 and 4.4.","section":"Figure 3"},{"comment":"The limitation paragraph says the method 'may cause a margin decline in ID class performance,' but the experiments show consistent ID improvements; please reconcile this statement with Tables 1-3 or remove it.","section":"§6"}],"recommendation":"major_revision","confidential_remarks":"The missing technical details (K value, exact loss terms, pseudo-label procedure, and OOD-mining analysis) are likely fixable in a revision and would materially improve the paper, so I recommend major revision rather than rejection. The authors should also report whether a held-out validation split was used when selecting hyperparameters, since several defaults are tuned on the same COCO-Open-CLS setting used for the headline results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a workmanlike extension of Unbiased Teacher to open-set semi-supervised detection, and the reported numbers are consistent and non-trivially better than the closed-set baseline. The paper deserves a serious referee, but it needs a revision that closes a real gap in the OOD training loop, not just polish.\n\nWhat is genuinely new is the particular combination: a supervised contrastive loss with a per-class memory pool (K+1 classes) plus OpenDet-style uncertainty classification, applied to the OSSOD setting. Each ingredient is known, but the recipe for OSSOD is new and, as the transfer-to-STAC experiment shows, it is portable. The experiments cover three benchmarks, the ablations make sense, and the qualitative figures are consistent with the tables. I believe the empirical claim that this helps both ID and OOD detection holds up in the reported settings.\n\nThe soft spots, in order of weight. First, the top-K background mining in Sec. 4.4 is the only OOD signal in the semi-supervised branch, yet the paper never defines K, never reports how often true OOD boxes land among the teacher's high-background proposals, and never measures recall@K. The stress-test note is right: if the teacher confidently assigns OOD objects to ID classes (its own Fig. 4 shows exactly this failure for the baseline), the mined set contains background, not OOD, and L_uc teaches the model to call background 'unknown'. The authors' own qualitative results show their method recovering OOD objects that UT mislabels, so the mechanism likely works sometimes, but the paper gives no evidence about when and how often. This is the load-bearing assumption and it is unexamined.\n\nSecond, the equations are incomplete. Eq. 3 drops the ROI classification loss from the supervised objective, which is a standard Faster R-CNN term; either it is an omission or the model is trained without it, and the reader cannot tell. Eq. 7's semi-supervised branch has only the OOD term, dropping ID classification entirely; that may be intentional to avoid pseudo-label noise, but the paper does not justify it beyond an ablation. Third, K aside, several hyperparameters (alpha_t, beta, q, dim, thresholds) are tuned on the same COCO-Open-CLS benchmarks used for the final tables. That is a circularity concern. No code is available, which compounds the issue.\n\nWho is this for? Anyone working on SSOD with open-set unlabeled data, or deploying detectors that need an 'unknown' output. It is a conference short paper, not a breakthrough, but the problem is real and the method is simple enough to build on. My recommendation: send it to review, but ask for a clarification of the OOD mining pipeline, the missing loss terms, and at least one analysis of recall of true OOD boxes among top-K background proposals. Without that, the main mechanism is unverified.","headline":"A simple, plausible OSSOD extension of Unbiased Teacher with consistent gains, but the OOD training signal rests on an unvalidated top-K background mining mechanism and key hyperparameters are undisclosed.","tokens_in":12765,"tokens_out":1874,"would_cite":false,"duration_ms":21592,"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":"This paper argues that two auxiliary losses can teach a semi-supervised object detector to label out-of-distribution objects as 'unknown' rather than forcing them into known classes.","keywords":["open-set semi-supervised object detection","out-of-distribution detection","contrastive learning","uncertainty classification loss","pseudo-labeling","feature memory pool","Faster R-CNN","teacher-student training"],"falsifier":"On the VOC-COCO setup, count how often objects from the 60 non-VOC classes appear among the teacher's top-K background proposals versus among proposals the teacher assigns to a known class above the pseudo-label confidence threshold; if OOD objects mostly fall in the latter group, the mined signal for $\\mathcal{L}_{uc}$ is empty and the reported $AP_u$ should not materialize. A second check: replace the top-K background mining with an equal number of randomly chosen background proposals; if $AP_u$ stays roughly the same, then the specific mining step is not what creates the unknown-class ability.","tokens_in":11678,"feed_emoji":"🎯","tokens_out":9997,"duration_ms":88398,"temperature":0.7,"pith_summary":"Open-set semi-supervised object detection means training a detector on a small labeled set of known classes while the unlabeled stream also contains objects from outside that set; standard semi-supervised detectors respond by forcing those outsiders into the closest known class. This paper argues the fix is to change the loss, not the architecture: a feature contrastive loss pulls same-class proposals together and pushes different classes apart, and an uncertainty classification loss teaches the detector to output a single 'unknown' label for out-of-distribution objects. The two losses are trained inside the ordinary teacher-student pseudo-labeling loop, so no separate out-of-distribution detector and no offline filtering stage is needed. The paper reports state-of-the-art results on COCO/VOC open-set benchmarks, including an unknown-class average precision of 9.82 on one setting, and shows the same two losses can be grafted onto another semi-supervised detector.","feed_headline":"A two-loss recipe gives semi-supervised detectors an 'unknown' class","feed_subtitle":"Feature clustering plus an uncertainty loss catches unknown objects (AP 9.82) without a second model.","key_machinery":"The mechanism is a pair of auxiliary losses bolted onto the standard teacher-student pseudo-labeling loop of a proposal-based detector. $\\mathcal{L}_{fc}$ is a supervised contrastive loss computed on 128-dimensional embeddings of RoI features: each proposal embedding is contrasted against a memory pool storing up to $q=256$ filtered embeddings per class, with only proposals passing $S_{IoU} > 0.7$ and $S_{cos} > 0.5$ written into the pool, and the unknown class stored without intra-class pull. $\\mathcal{L}_{uc}$ is a weighted uncertainty classification loss: for known classes it is ordinary weighted cross-entropy, while for the unknown class it computes softmax over background plus unknown only and reweights the unknown term by $(1-p_k)^\\alpha p_k$, where $p_k$ is the summed probability of known classes. The top-K highest background-class proposals supply the pseudo-ground-truth OOD signal during semi-supervised training. Together the two losses make the existing classifier output an 'unknown' category without changing the detector architecture.","core_discovery":"The central discovery is that a closed-set semi-supervised detector can be turned into an open-set detector by joint optimization at two levels of the same model. At the feature level, the loss $\\mathcal{L}_{fc}$ forms class clusters in the embedding space using a memory pool of per-class region embeddings filtered by IoU and cosine similarity; because out-of-distribution objects are semantically varied, it deliberately avoids compacting the OOD class and only separates it from known classes. At the logits level, the loss $\\mathcal{L}_{uc}$ mines candidate OOD instances by taking the top-K background-class proposals and forces the classifier to assign probability mass to a unified class $K+1$ ('unknown'), with the OOD term weighted by uncertainty $(1-p_k)^\\alpha p_k$. Trained in a teacher-student semi-supervised pipeline, the resulting detector reports both known classes and the unknown class; on the COCO/VOC benchmarks the paper reports $mAP_k$ up to 25.05 and $AP_u$ up to 9.82, and ablations attribute the unknown-class ability mainly to $\\mathcal{L}_{uc}$.","pith_inferences":["The paper does not analyze how often OOD objects actually land in the teacher's top-K background proposals; a natural extension would be to measure that hit rate and make K adaptive, since the entire OOD training signal depends on this mine.","Because the losses attach to shared RoI features and logits, the same recipe should transfer to open-set versions of anchor-free and transformer-based detectors, though the paper only demonstrates transfer to one other SSOD method.","The acknowledged trade-off between ID accuracy and unknown accuracy implies a Pareto frontier; the paper sweeps only the loss weight $\\beta$, and a systematic multi-objective study of thresholds and temperatures could find better operating points."],"forward_implications":["If the claim holds, any closed-set pseudo-label-based semi-supervised detector can be upgraded to open-set operation by adding these two losses, with no second model and no extra fine-tuning stage.","Detectors trained this way would stop presenting novel objects as high-confidence known classes and would instead emit a single 'unknown' label, a property directly relevant to autonomous vehicles and open-world perception systems.","The per-class memory pool gives the model a growing set of class prototypes during training, which should strengthen representation learning for known classes as well as separating unknowns.","The method remains effective with as few as 1,000 labeled images (unknown AP 4.25 on the 20/60 setting), suggesting the OOD signal from background mining is usable even when supervision is scarce."],"supporting_citations":[{"why":"Provides the teacher-student pseudo-labeling pipeline and the closed-set baseline this method extends.","marker":"[23]"},{"why":"Supplies the uncertainty-estimation idea that the logits-level loss adapts for OOD classification.","marker":"[7]"},{"why":"Defines the open-set SSOD task and the offline OOD-detector baseline the paper argues is insufficient.","marker":"[22]"},{"why":"Is the recent OSSOD comparison method whose threshold-free OOD head the paper aims to beat.","marker":"[33]"},{"why":"Is the supervised-contrastive formulation on which the feature-level loss is built.","marker":"[15]"},{"why":"Is the semi-supervised detector used in the transfer experiment that shows the two losses are plug-and-play.","marker":"[27]"},{"why":"Is an open-world detection baseline compared in the tables and the target of the paper's critique of probability-only OOD learning.","marker":"[14]"},{"why":"Is the Faster R-CNN detector architecture used for region proposals and classification heads.","marker":"[5]"},{"why":"Supplies the feature pyramid from which RoIAlign features are extracted for the contrastive embeddings.","marker":"[20]"}],"fun_headline_variants":["Two losses teach semi-supervised detectors to spot unknowns","Open-set detection from clustering and uncertainty losses","Feature-logits contrastive learning opens up SSOD","Semi-supervised detector learns to flag unknown objects","Two-level loss adds an 'unknown' class to SSOD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on the teacher detector surfacing out-of-distribution objects as high-scoring background proposals, because the uncertainty loss mines its unknown-object training signal from the top-K background boxes; if the teacher confidently assigns such objects to a known class, the loss never sees them as unknown.","fun_headline_variants_meta":{"raw":{"variants":["Two losses teach semi-supervised detectors to spot unknowns","Open-set detection from clustering and uncertainty losses","Feature-logits contrastive learning opens up SSOD","Semi-supervised detector learns to flag unknown objects","Two-level loss adds an 'unknown' class to SSOD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000502,"raw_usage":{"total_tokens":2451,"prompt_tokens":938,"completion_tokens":1513,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":1437}},"tokens_in":554,"tokens_out":1513,"duration_ms":12469,"temperature":1.0,"reasoning_tokens":1437,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:56:55.487973+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the VOC-COCO setup, count how often objects from the 60 non-VOC classes appear among the teacher's top-K background proposals versus among proposals the teacher assigns to a known class above the pseudo-label confidence threshold; if OOD objects mostly fall in the latter group, the mined signal for $\\mathcal{L}_{uc}$ is empty and the reported $AP_u$ should not materialize. A second check: replace the top-K background mining with an equal number of randomly chosen background proposals; if $AP_u$ stays roughly the same, then the specific mining step is not what creates the unknown-class ability.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the teacher-student pseudo-labeling pipeline and the closed-set baseline this method extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the uncertainty-estimation idea that the logits-level loss adapts for OOD classification."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the open-set SSOD task and the offline OOD-detector baseline the paper argues is insufficient."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the recent OSSOD comparison method whose threshold-free OOD head the paper aims to beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is an open-world detection baseline compared in the tables and the target of the paper's critique of probability-only OOD learning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the Faster R-CNN detector architecture used for region proposals and classification heads."}],"review_version":1}