acceptodds
Under review as a conference paper at ICLR 2027

Multi-Label Generalized Category Discovery

Abstract

In this paper, we introduce Multi-Label Generalized Category Discovery (ML-GCD), a new setting for jointly recognising Known categories and discovering Novel categories in multi-label data. It extends the single-label formulation of Generalized Category Discovery (GCD) to category-wise presence estimation. In ML-GCD, labelled instances contain only Known categories, whereas the unlabelled pool spans Known-only, Novel-only, and mixed instances with all annotations withheld. For unlabelled instances, cross-view category learning derives supervision from model-generated targets, but the resulting consistency objective lacks an independent, fixed reference for representation learning. Our key idea is to couple category discovery with the preservation of global and local visual structure, using a fixed pretrained visual reference independent of evolving category assignments. We propose Patch-guided Discovery (PaD), which transfers Known-category knowledge to unlabelled instances, preserves global feature relations, and retains local visual information. Experiments on VOC2007, MIRFlickr, and COCO2014 demonstrate the effectiveness of PaD for ML-GCD, with ablation studies supporting the value of fixed visual guidance in complementing model-generated supervision.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Category Discovery, Semi-superviesd Learning, Transfer Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ieQJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies Multi-Label Generalized Category Discovery (ML-GCD), where labeled training images contain only known categories, while unlabeled images may contain novel categories, or both known and novel categories. The goal is to recognize known categories and discover consistent novel-category identities across images, assuming the number of novel categories is given. The proposed Patch-guided Discovery (PaD) framework uses a DINO-pretrained Vision Transformer with independent sigmoid outputs and sparse cross-view pseudo-labels. A frozen teacher guides student learning through known-category prediction distillation, global relation preservation, and spatially aligned patch distillation. Experiments on VOC2007, MIRFlickr, and COCO2014 include baseline comparisons, component ablations, and analyses of mixed-category images. Anonymous novel-category outputs are globally matched to ground-truth categories for evaluation.

Critical strengths
  1. The paper introduces a valuable task setting and provides a detailed discussion of relevant prior work. To the best of my knowledge, the most closely related studies are appropriately acknowledged.
  2. The paper presents a distinctive methodological design and describes its components and implementation in detail.
  3. Experiments on three datasets evaluate performance using multiple metrics, while component ablations provide evidence for the contributions of individual modules.
  4. The paper transparently reports negative results and experimental limitations, including reduced joint recognition on mixed-category images, variability across random seeds, and optimization issues in some baseline adaptations.
Critical weaknesses

1. Weak exposition and insufficient motivation for key design choices.

The paper devotes substantial space to equations and implementation details but does not clearly connect the problem’s challenges to the proposed design choices. For example:

  • Although the differences between ML-GCD and earlier settings are specified, its motivation remains unclear: which applications require these supervision restrictions, and what distinct challenges do they introduce beyond similar weakly supervised settings? Later sections do not resolve these questions.
  • The method description is highly technical but offers limited justification for using DINO, applying Sinkhorn normalization, assigning a second positive pseudo-label to half of the unlabeled images, or choosing the associated hyperparameters.

Conversely, the task definition, global evaluation matching, and baseline adaptations are explained too briefly. For example:

  • Section 2 describes DLCL and MCUL through the phrases “latent label columns” and “latent columns with missing observed entries.” These descriptions do not adequately convey their supervision assumptions to readers unfamiliar with the papers.
  • The evaluation protocol is difficult to follow even for readers familiar with Average Precision. The equation is explicit, but the meaning of “global” alignment—one matching computed over the entire evaluation set and reused across subsets—must largely be inferred.

These issues hinder both comprehension and assessment of the paper’s contributions.

2. Insufficient justification of the task setting and its distinction from prior work.

ML-GCD requires all training images containing novel categories to enter the unlabeled pool, with all annotations—including known-category annotations—withheld. The paper does not explain why it excludes the natural alternative of retaining known-category annotations in such images, or which applications motivate this restriction.

Although DLCL and MCUL are acknowledged, their supervision assumptions are not explained sufficiently to establish the significance of the differences. Table 1 also risks conflating task specifications with method capabilities. These studies do not stipulate “Labelled subset excludes Novel” or “Fully unlabeled training pool,” but this does not establish that their methods cannot accommodate these conditions through adaptation.

3. Insufficient explanation of the discovery mechanism and the main performance gains.

The method uses anonymous category slots, sparse cross-view pseudo-labels, and self-training, but provides limited explanation of how these components encourage the slots to represent the intended semantic categories. The initialization of category directions and the rationale for pseudo-label selection are also insufficiently specified.

On VOC, Novel mAP rises from 6.80 for ML-SimGCD to 53.95 for the category-learning control, then to 54.93 with PaD. Thus, the largest gain occurs before the distillation components emphasized in the paper are added. The ablations mainly isolate these distillation components rather than explain the much larger gain from the base discovery framework. It remains unclear how much improvement comes from the loss, pseudo-label rules, the samples receiving supervision, or baseline adaptation.

4. Insufficiently controlled external baseline comparisons.

Table 3 mixes adapted external methods, an internal ablation, and a hybrid framework without clearly distinguishing their purposes. The category-learning control is a direct ablation; MulSupCon with an adapter combines an existing representation-learning framework with a discovery module; and ML-SimGCD is an adapted discovery method whose particularly low performance warrants investigation. On MIRFlickr and COCO, the main table compares PaD only with the internal control.

The appendix includes relevant methods such as DLCL, MCUL, and DMNL. However, these methods use frozen features while PaD fine-tunes the backbone, and some adaptations have unresolved convergence or calibration issues. These differences limit the conclusions that can be drawn about PaD’s advantage over prior discovery methods.

5. Insufficient analysis of discovery quality and performance trade-offs.

The component ablations and mixed-image evaluations are useful, but several central questions remain unanswered:

  • What do the novel slots learn? Do they duplicate known categories, split individual categories, capture background patterns, or fail to cover some categories?
  • Are the proposed relation and patch constraints necessary, or could freezing the backbone or applying a simpler feature-matching loss achieve comparable results?
  • Why do aggregate mAP, F1, or precision improve while joint known–novel recognition declines on VOC and COCO? Is this caused by score calibration or by reduced recognition of one category group?

These questions directly concern the stated goal of recognizing known and novel categories when they co-occur. The current analysis does not sufficiently explain how the method achieves this goal or where it fails.

Questions for authors
  1. What accounts for the large gain from ML-SimGCD to the category-learning control? On VOC, Novel mAP increases from 6.80 to 53.95 before distillation is added. Can the authors identify which changes drive this improvement and provide evidence that the ML-SimGCD adaptation was appropriately configured?
  2. What evidence establishes PaD’s advantage over the closest prior methods? The appendix reports DLCL, MCUL, and DMNL adaptations, but their frozen representations and, in some cases, unresolved optimization issues complicate comparison. Are there results under comparable representation-training conditions, or evidence that these adaptations were adequately optimized?
  3. How are the novel classifier directions initialized, and what evidence shows that the resulting slots capture distinct target categories? In particular, do the reported results reveal duplicated semantics, overlap with known categories, or novel categories that remain undiscovered?
2 comments
AuthorsOpenReview ↗

Response to Reviewer ieQJ

Thank you. We appreciate your recognition that we report negative results and limitations transparently; this was deliberate. Section and table numbers refer to the revised manuscript.

1. Setting, motivation and exposition.

ML-GCD deliberately uses a strict supervision boundary to test whether annotations of known-category compositions transfer to unannotated compositions containing novel categories (Sec. 3). Retaining Known annotations in mixed images is a relaxation with additional supervision; it must be compared within its own regime (Appendix A), and the revised text positions it as an extension. Preliminaries now state the visible labels and the supplied Novel-category count.

Table 1 marks conditions not stipulated in the original formulations, not inability to adapt, and we adapted DLCL, MCUL and DMNL to ML-GCD supervision (Appendices G.1, I.8). We will replace the compressed descriptions of DLCL and MCUL in Related Work with explicit statements of which category labels are known and which annotations are missing.

On design choices (Sec. 4): DINO follows GCD's self-supervised pretraining without manual category labels. Sinkhorn supplies a batch-balancing prior for ranking targets, not multi-label probabilities. Selecting a second positive for a prescribed fraction of images is a sparse training prior (about 1.5 positives per image), not an estimate of true label counts; Table 11 shows non-monotonic sensitivity to this fraction and to the Sinkhorn iteration count. We will move component purposes ahead of the equations and consolidate numerical settings in the appendix. Global alignment is one AP-maximising Novel matching over the full test set, reused across metrics and subsets, with Known correspondences fixed (Sec. 5.1, Appendix A).

2. Attribution and the ML-SimGCD adaptation (Q1).

The largest gain (VOC Novel mAP 6.80 to 53.95) comes from multi-label target adaptation, which we regard as a core part of our contribution; the revised text foregrounds it. Table 3 separates the steps cumulatively: cross-view learning restricted to unlabelled images (6.80 to 10.42), ASL on Known labels (8.69), sparse masking on unlabelled images (32.60) and ASL on the selected entries (53.95). PaD then reaches 54.93. The largest steps coincide with replacing dense targets by sparse ones, which admit multiple positives and leave intermediate entries unsupervised instead of treating every unselected category as absent, and with ASL on those entries. These are ordered contrasts, not independent causal estimates.

ML-SimGCD applies rescaled Bernoulli targets and BCE to both labelled and unlabelled images (Appendix A). It shares the backbone, trainable block, initial head, learning rate, batch size and 200-epoch horizon with the other single-stage methods.

3. External comparisons and simpler preservation (Q2).

Table 2 separates external adaptations, the internal learner and the hybrid MulSupCon framework; Table 14 reports the longer 150+200 MulSupCon schedule, which neither establishes convergence nor aligns total budgets. DLCL, MCUL and DMNL use frozen features and therefore do not isolate discovery objectives from representation learning. We state this in the main text, do not claim superiority over the original methods, and leave fine-tuned counterparts for future work. For simpler preservation, Table 10 compares a frozen backbone, CLS matching (Novel mAP 56.32) and an EMA teacher (55.74) with PaD (54.93). These controls are competitive, and we do not claim that every PaD component is necessary.

4. Discovery quality and mixed-image recognition (Q3).

Classifier directions are randomly initialised and normalised, and no Novel labels determine them (Appendix B.1). Shared outputs and cross-view supervision encourage consistent predictions across views but do not guarantee distinct semantic categories. In the additional VOC pair (Appendix E.1), ten Novel categories select only seven distinct highest-AP raw outputs in both models, despite PaD's higher assigned Novel mAP, which shows overlap in preferred outputs. We will add image-grounded inspection of category overlap and missed categories.

On joint recognition, mAP and aggregate F1 do not require both category groups to be correct within each image, whereas joint hit does. The VOC/COCO joint-hit losses accompany lower Known recall even though COCO Novel recall improves (Table 12), and the revised conclusion states this deficit directly. Whether it reflects score calibration or reduced recognition of one category group is not yet resolved. The Known-only recall drop in Appendix E.1 suggests it is not confined to mixed images, and we will add teacher-calibration measurements on Known-only versus mixed images.

Thank you for taking the review comments seriously and investing substantial effort in providing additional explanations and experiments. I appreciate these efforts. However, several key concerns remain unresolved, and I therefore maintain my original rating.

W1. Weak exposition and insufficient motivation for key design choices.

I acknowledge the additional explanations and experiments. Nevertheless, the writing issues in the paper continue to substantially affect its readability and the completeness of its arguments. Technical details are organized within a insufficiently clear logical framework, making it difficult for future readers to understand the connections between the underlying challenges, the proposed design choices, and their intended effects. Furthermore, explanations that are essential to understanding the paper's core contributions, should not rely primarily on the rebuttal or supplementary material. Therefore, until the main paper undergoes substantial revision to incorporate these explanations and establish a clearer line of argument, I will maintain this concern unresolved.

W2. Insufficient justification of the task setting and its distinction from prior work.

I still think that the proposed ML-GCD setting is quite similar to existing task formulations (table 1), and that its substantive distinctions from them have not been sufficiently established.

I do not dispute that these constraints define a different supervision protocol. However, the authors have not sufficiently explained why these differences warrant an independently motivated research task. In particular, it remains unclear which practical applications necessitate these restrictions and what fundamentally different research challenges they introduce compared with existing settings. Consequently, I believe that the novelty of the proposed formulation, particularly its claim of being the first to formulate Multi-Label Generalized Category Discovery, is limited.

W4. Insufficiently controlled external baseline comparisons.

Given the substantial similarity between ML-GCD and related settings, I think the effective comparisons for these related methods, including DMNL, DLCL, and MCUL, is particularly important. However, the main paper does not provide convincing direct comparisons, and the adaptation experiments in the supplementary material do not adequately address this concern.

In addition, these experiments suffer from substantial comparability issues. The related methods use frozen features, whereas PaD fine-tunes the backbone. Furthermore, some adaptations exhibit unresolved optimization convergence, kernel degeneracy, or score-calibration issues. These factors make it difficult to assess the relative capabilities of the methods themselves. In particular, when the adaptations of certain baselines have not been sufficiently validated, their extremely poor performance cannot serve as reliable evidence of PaD's superiority. I appreciate the authors' transparency in reporting these experimental limitations, but this does not resolve the underlying concern regarding the validity of the comparisons.

W5 & Q2. Insufficient analysis of discovery quality and performance trade-offs.

I still find the current experimental evidence insufficient to establish confidence in the proposed method.

First, the actual quality of Novel category discovery remains insufficiently analyzed. The additional experiments reported by the authors show that the 10 Novel categories correspond to only 7 distinct raw outputs when selecting the highest-AP output for each category. This raises concerns about whether different Novel slots actually learn semantically distinct categories. Furthermore, the decline in joint Known–Novel recognition on VOC and COCO remains insufficiently explained. Given that recognizing Known and Novel categories simultaneously is a central objective of the proposed task, this issue should not be treated merely as a secondary performance trade-off.

Second, the methodological contribution of PaD remains unclear. The performance gains brought by the proposed global-relation preservation and local-feature preservation constraints are limited, and the additional comparisons with simpler alternatives further weaken the case for their necessity. For example, CLS matching and the EMA patch teacher achieve Novel mAP scores of 56.32 and 55.74 on VOC, respectively, both exceeding PaD's 54.93. The authors also acknowledge that these simpler alternatives are competitive. Therefore, the current experiments have not established a clear advantage of the proposed complex constraints over simpler feature-preservation strategies.

Reviewer_3ARROpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces Multi-Label Generalized Category Discovery (ML-GCD) and proposes Patch-guided Discovery (PaD), which combines sparse cross-view pseudo-label learning with known-category distillation, global relation preservation, and patch-level feature alignment.

Critical strengths

The ML-GCD setting is interesting.

The paper provides component ablations and evaluates performance across three multi-label datasets.

Critical weaknesses
  1. Poor table readability.

  2. The proposed method primarily combines established techniques, including knowledge distillation, relational feature preservation, patch-level alignment, and pseudo-label learning. The paper does not sufficiently demonstrate what fundamentally new methodological insight PaD contributes beyond integrating these existing components.

  3. A central motivation of ML-GCD is to recognize Known and Novel categories simultaneously within the same image. However, Table 5 shows that PaD substantially reduces the joint hit rate on VOC2007 (27.36% to 13.08%) and COCO2014 (40.52% to 27.53%).

  4. Compared with the category-learning control, PaD improves Novel mAP by only 0.98 points on VOC2007 and 1.89 points on COCO2014. Furthermore, the paper reports that Novel mAP improves in only two out of three paired seeds on each dataset. These results provide limited evidence of consistent improvements in novel category discovery.

  5. Table 4 shows that removing Known-prediction distillation actually improves Novel mAP from 54.93 to 56.51. The contribution of patch distillation is also variable across seeds. These findings suggest that the proposed components may introduce conflicting optimization objectives, which are not adequately investigated.

Questions for authors
  1. How do the authors explain the substantial degradation in joint Known–Novel recognition on VOC2007 and COCO2014, given that simultaneous recognition is a central motivation of ML-GCD?

  2. Why does removing Known-prediction distillation improve Novel mAP? Does this indicate a conflict between preserving Known-category knowledge and discovering Novel categories?

  3. Can the authors provide stronger evidence that the proposed combination of distillation objectives offers benefits beyond existing representation-preservation techniques?

  4. Can the authors provide more comprehensive baseline comparisons and additional experiments to establish the robustness of the reported improvements?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 3ARR

Thank you for identifying the weaknesses in presentation, contribution and robustness. We have revised the main text to separate target adaptation, component effects and mixed-image recognition. Section and table numbers refer to the revised manuscript.

1. Contribution and presentation.

PaD builds on existing operations, and Related Work says so. The contribution lies in three places: (i) the ML-GCD formulation and evaluation protocol; (ii) a multi-label target construction (sparse cross-view targets with ASL) that accounts for most of the gain in our sequential study (VOC Novel mAP 6.80 to 53.95, Table 3); and (iii) a controlled comparison showing that Known-prediction, global-relation and local-feature preservation affect Known recognition and Novel discovery differently, and that improved ranking does not guarantee joint recognition. Preliminaries now state the co-occurrence challenge and supervision regime, and Method separates the base learner's targets from PaD's preservation constraints. On readability, Table 2 separates training routes and Table 4 presents the eight component combinations as a checkmark matrix; we will replace the remaining abbreviated labels with descriptive names.

2. Component interactions and alternatives (Q2, Q3).

Q2: yes, this indicates a trade-off between preserving Known predictions and discovering Novel categories. In the full-model removal comparison (Table 4), Known distillation adds 8.71 Known points but reduces Novel mAP by 1.58, while relation and patch constraints add 1.34 and 1.32 Novel points. Full PaD has the highest All/Known means, whereas relation preservation alone has the highest Novel mean. The endpoint gradient cosines between Known distillation and category learning (−0.916-0.916 on VOC, −0.934-0.934 on COCO; Appendix D) are consistent with conflicting update directions, though they describe the observed endpoints rather than the whole trajectory. The main text now discusses this interaction.

Q3: Table 10 compares PaD with simpler alternatives: CLS matching (Novel mAP 56.32), an EMA patch teacher (55.74) and a frozen backbone, against 54.93 for PaD. These alternatives are competitive on Novel mAP. Our claim concerns the recognition/discovery balance when knowledge is preserved, not exclusivity of the fixed patch teacher, and the main text now says so.

3. Mixed-image recognition (Q1).

We agree that this result sits uneasily with the task motivation; the revised Sec. 5.6 and the conclusion state it directly. Joint hit improves from 37.10 to 48.53 on MIR but falls from 27.36 to 13.08 on VOC and from 40.52 to 27.53 on COCO. The declines accompany Known-recall drops (76.84 to 52.30; 56.97 to 23.83), while COCO Novel recall rises from 48.20 to 51.82. This identifies a Known-recognition deficit, not improved simultaneous recognition. One hypothesis is that the Known teacher head, fitted only on Known-only images, is miscalibrated on mixed images. We will add teacher-score and calibration measurements on Known-only versus mixed images to test it.

4. Robustness and external comparisons (Q4).

Seven complete MIR pairs give mean All/Novel gains of 4.41/4.00 points, positive in seven and six pairs. VOC's original eight pairs give −0.61/−0.07-0.61/-0.07, and five additional pairs give +2.21/+4.89+2.21/+4.89. The MIR unadjusted sign-flip p=0.0156p=0.0156 is the smallest value attainable with seven pairs, but no comparison passes Holm adjustment across the 12 exploratory tests (Appendix C). We present MIR's repeatedly positive gains and VOC's cohort dependence separately.

For external comparisons, Table 2 separates single-stage adaptations, the internal learner and the MulSupCon two-stage framework, and Table 14 gives complete three-seed VOC/MIR results for a longer 150+200 MulSupCon schedule. DLCL, MCUL and DMNL are evaluated on frozen features (Appendices G.1, I.8). We label them as such: they do not isolate discovery objectives from representation learning, and we do not claim superiority over the original methods. Fine-tuned counterparts are left for future work.

Reviewer_oZpJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper formulates a new problem, ML-GCD, a multi-label extension of GCD in which labelled images contain only Known classes, and proposes PaD. PaD regularises a sparse Sinkhorn pseudo-labelling learner with three distillation terms from a frozen DINO teacher. PaD outperforms the evaluated baselines on average, but with high variance across runs.

Critical strengths
  • The core contribution of this paper is the formulation of a new task, ML-GCD. The supervision boundary of ML-GCD is formally defined. Table 1 contrasts it with DMNL, DLCL and MCUL, and the evaluation protocol covers ranking, set prediction and composition-specific metrics. The setting is a natural and useful benchmark direction.
Critical weaknesses
  • High variance relative to the gains. In Table 3, PaD’s VOC All mAP has a standard deviation of 5.35, while its mean improvement over the category-learning control is only 4.04. With variance this high, it is hard to conclude that the proposed method is better. The authors should increase the number of training seeds and report significance tests.
  • Results on mixed images contradict the motivation. The paper is motivated by images containing both Known and Novel objects, yet Table 5 shows that PaD lowers joint hit on mixed images from 27.36 to 13.08 on VOC and from 40.52 to 27.53 on COCO. The reported precision gain may simply reflect PaD predicting fewer labels, so it is not evidence of better recognition. The authors should explain this decline, for example whether the Known teacher head, trained only on Known-only images, is miscalibrated on mixed images.
Questions for authors
  • Can you report per-seed results for Table 3 and explain the high standard deviations in some results, for example, whether one seed collapsed or performed abnormally well?
  • With at least 5 seeds and a paired significance test, does PaD’s improvement over the category-learning control remain significant on all three datasets?
  • Why does PaD lower joint hit on VOC and COCO mixed images? Is the Known teacher head, trained only on Known-only images, miscalibrated on mixed images, e.g., does it give lower Known probabilities when Novel objects are present? More in-depth analysis is desired.
  • Why did the authors use only DINOv1 as the backbone instead of DINOv2/v3 and their larger variants? Do the gains still hold with these stronger backbones?
1 comment
AuthorsOpenReview ↗

Response to oZpJ

Thank you for the careful review. Your two main concerns, the reliability of the VOC gains and the drop in mixed-image joint recognition, are well founded, and we have revised the paper to address them directly. Section and table numbers refer to the revised manuscript.

1. Variance and significance (Q1, Q2).

Sec. 5.5 and Appendix C (Tables 8–9) now report 13 complete VOC pairs (seeds 0–7 and 8–12, as separate cohorts) and seven complete MIR pairs, with mean paired differences, 95% paired-tt intervals, exact sign-flip tests and Holm adjustment over 12 comparisons. MIR seed 5 failed numerically in both arms and on retry; it is excluded without replacement or imputation. We will add a per-seed table to the appendix.

Q2: under the current data, the improvement is not statistically significant on all three datasets. On MIR, the mean All/Novel gains are +4.41/+4.00+4.41/+4.00 points, positive in 7/7 and 6/7 pairs (All interval [3.42, 5.40]). The unadjusted sign-flip p=0.0156p=0.0156 is the smallest value attainable with seven pairs; Holm adjustment over the 12 exploratory comparisons gives 0.1875. We therefore describe the MIR gain as consistent in sign and magnitude but not confirmatory. On VOC, All/Novel differences are −0.61/−0.07-0.61/-0.07 (seeds 0–7) and +2.21/+4.89+2.21/+4.89 (seeds 8–12), with All intervals including zero in both cohorts. COCO retains its three original pairs (+1.22/+1.89+1.22/+1.89). The revised text no longer presents the original VOC mean gain as stable and limits its claims to the MIR ranking gain and the original-panel set-prediction results.

Q1: in the original VOC panel (Table 2), the large All SD (5.35) comes mainly from Known mAP (SD 10.10); PaD's Novel SD (0.73) is below the base learner's (2.23). Across the 13 VOC pairs, the Novel difference is positive in ten, versus five for All and four for Known, so the instability is concentrated in Known performance. We have not yet determined whether a single abnormal seed drives it. We will add run-level learning curves and prediction distributions to answer this, and complete the planned COCO paired extension.

2. Mixed-image recognition and calibration (Q3).

We agree that the VOC/COCO results limit any claim of improved simultaneous recognition. Sec. 5.6 and Appendix E now state this as a finding: mAP measures category-wise ranking and F1 is aggregate, whereas joint hit requires a correct Known and a correct Novel prediction in the same mixed image.

Joint hit falls from 27.36 to 13.08 on VOC and from 40.52 to 27.53 on COCO. These declines accompany Known-recall drops from 76.84 to 52.30 and from 56.97 to 23.83, while Novel recall is roughly maintained (39.90 to 37.03; 48.20 to 51.82). On MIR, joint hit improves from 37.10 to 48.53 with approximately unchanged Known recall. Your fewer-labels explanation is consistent with the additional VOC pair (Appendix E.1), where the mean predicted label count falls from 3.85 to 2.65 alongside lower Known recall; we will report these counts in Appendix E.1. We therefore no longer treat precision or F1 as evidence of better joint recognition.

On calibration: the teacher head is fitted on Known-only images but supervises Known outputs in labelled and unlabelled batches (Sec. 4.3), so your hypothesis is plausible and not yet directly tested. Two observations are consistent with a Known-distillation effect. Removing it returns VOC F1 and label-count MAE to base-learner levels (Table 13), and its endpoint gradient cosine with category learning is −0.916-0.916 (VOC) and −0.934-0.934 (COCO) (Appendix D). However, in the additional VOC pair, class-standardised Known recall also falls on Known-only images (95.98 to 83.05) with lower Known mAP (65.34 to 59.01), so the deficit is not confined to mixed images at one threshold. We will add teacher Known scores on Known-only versus mixed images, and joint hit and Known recall for the Table 4 variants.

3. Backbone choice (Q4).

Sec. 4 states that DINO-v1 follows GCD's self-supervised pretraining setup without manual category labels, and all methods in Table 2 share this initialisation. Appendix F specifies a DINOv2 ViT-B/14 configuration, but the two complete MIR pairs available so far are not enough for conclusions across datasets and seeds. We will complete paired DINOv2 evaluations on VOC and MIR with the same statistical and mixed-image analyses, and report the outcomes whichever way they go. Larger backbones and DINOv3 remain untested.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.