acceptodds
Desk rejected at ICLR 2027

DURA: Dual Stage Unknown Aware Representation Adaptation for Underwater On-the-Fly Category Discovery

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

On-the-Fly Category Discovery (OCD) aims to recognize known categories and discover novel ones from an unlabeled data stream. In underwater scenarios, fine-grained inter-class similarity and substantial appearance variations make this task particularly challenging. Existing methods either keep offline representations fixed during inference or adapt representations learned mainly from known categories. To address this issue, we propose DURA, a Dual Stage Unknown Aware Representation Adaptation framework for underwater OCD. In the offline training stage, DURA constructs pseudo unknown samples and explicitly separates them from known category prototypes, preparing the representation for novel category discovery. In the online inference stage, DURA dynamically expands the prototype memory and adapts both the encoder and category prototypes using the observed query stream, allowing the model to adjust to emerging category structures. For systematic evaluation, we establish UOCDBench, to our knowledge the first benchmark dedicated to underwater OCD. It combines our newly introduced AquaOCD343, containing 343 categories and 67,651 images, with four existing underwater datasets. Experiments on UOCDBench and standard OCD benchmarks demonstrate that DURA consistently achieves state-of-the-art performance, particularly improving novel category discovery while maintaining strong known category recognition.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Underwater Visual Recognition, On-the-Fly Category Discovery, Unknown Aware Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_UT7MOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose DURA, a dual-stage unknown-aware representation adaptation framework for On-the-Fly Category Discovery (OCD). The method operates in two stages: first, it generates pseudo-unknown samples during offline training and uses maximum-similarity repulsion to separate them from known-category prototypes; second, it adapts both the encoder and prototype memory online as unlabeled query samples arrive, enabling the model to recognize known categories and discover new ones. To support systematic evaluation in underwater scenarios, the authors introduce AquaOCD343 and construct UOCDBench by combining it with four existing underwater datasets. Experiments on UOCDBench and standard OCD benchmarks demonstrate consistent improvements over recent state-of-the-art methods, particularly in novel-category discovery.

Critical strengths
  1. The paper addresses online category discovery in underwater environments, where known and previously unseen categories appear in a query stream. It introduces DURA, a two-stage framework that separates pseudo-unknown samples from known prototypes during offline training and updates the encoder and prototype memory during online inference.
  2. The DURA framework is clearly structured, with separate offline and online objectives. The offline stage combines angular-margin classification with pseudo-unknown separation, while the online stage uses entropy minimization, prototype alignment, inter-class separation, and confidence- and support-aware prototype refinement. The method is also evaluated with CLIP, DINO, and DINOv2 backbones.
  3. The paper evaluates DURA on UOCDBench, which includes AquaOCD343 and four existing underwater datasets, as well as five standard OCD datasets. It reports Old, New, and All accuracy under Greedy-Hungarian and Strict-Hungarian protocols. Ablation studies examine the main components, and category-count analysis provides additional information about the number of discovered categories.
Critical weaknesses
  1. As the number of categories discovered online increases, each query must be matched against more prototypes, which may increase the computational and memory costs of the online adaptation stage. The paper currently lacks an analysis of inference latency, throughput, and memory usage under different numbers of prototypes. Relevant experiments are needed to evaluate the feasibility of the method in long-term real-time operation.
  2. DURA decides whether to create a new prototype based on the similarity between a sample and the existing prototypes. When samples from the same class have large visual differences, some of them may be incorrectly treated as new categories, causing the same real class to be split into multiple prototypes. Why are strategies such as prototype merging not considered to mitigate this issue?
  3. During the Online Inference stage, the model uses its own predictions for representation adaptation and prototype updates. Therefore, incorrect sample assignments may affect the prototypes and further influence subsequent predictions. Reporting the proportion of incorrect assignments and comparing the similarity between the prototypes and the samples of the corresponding classes before and after the updates would provide a clearer view of the impact of online updating.
Questions for authors

Please see the weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Weakness 1

We sincerely appreciate the reviewer’s constructive comments on computational scalability, category fragmentation, and erroneous online feedback. We have completed the experiments reported below. After completing and analyzing the remaining additional experiments requested by every reviewer, we will incorporate the results into the revised manuscript’s appendix and clearly highlight all revisions in blue.

Response to Weakness 1

We appreciate the reviewer’s suggestion to analyze computational costs as the prototype memory grows. Our additional measurements show that DURA maintains efficient online processing with limited memory growth across the evaluated complete query streams. At the larger observed prototype counts, processing time remains approximately 1.1 ms per image, with throughput around 900 images/s.

We evaluate DURA-DINOv2 on AquaOV255 and AquaOCD343 using an NVIDIA RTX 5090 and a query batch size of 64, with five random query permutations per dataset. Each run starts from the same offline checkpoint and initial state. Prototype counts grow naturally under the original discovery and adaptation procedures.

The measured processing time includes input transfer, feature extraction, sequential prototype matching, new-prototype creation, online encoder adaptation, and prototype refinement. CUDA synchronization is performed at the timing boundaries. Data loading and final Hungarian evaluation are excluded.

Table Response1. Online processing costs at different naturally occurring prototype counts. Results are mean ± sample standard deviation across runs.

DatasetPrototype count before a batchValid runsAmortized processing time (ms/image)Throughput (images/s)Peak GPU memory (MiB)
AquaOV255K < 25051.446 ± 0.135696.44 ± 66.861800.62 ± 0.28
AquaOV255250 ≤ K < 50051.115 ± 0.053898.47 ± 40.061803.67 ± 0.05
AquaOV255500 ≤ K < 75051.119 ± 0.045895.12 ± 34.731804.73 ± 0.15
AquaOV255750 ≤ K < 1,00041.113 ± 0.058900.07 ± 44.811804.20 ± 0.69
AquaOCD343K < 25051.256 ± 0.043796.94 ± 27.241800.81 ± 0.35
AquaOCD343250 ≤ K < 50051.138 ± 0.035879.25 ± 26.771803.88 ± 0.07
AquaOCD343500 ≤ K < 75051.102 ± 0.034908.46 ± 27.641803.54 ± 0.54

The amortized processing time is calculated by dividing the total batch processing time by the number of processed queries. Peak memory is the maximum PyTorch allocated GPU memory within each interval, calculated per run and then summarized across runs. Four AquaOV255 runs reach the highest interval.

From the lowest to the highest prototype-count intervals, peak GPU memory increases by 3.58 MiB on AquaOV255 and 2.73 MiB on AquaOCD343. Average persistent prototype storage in the highest intervals is 2.25 MiB and 1.62 MiB, respectively. Together with the processing-time measurements, these results strengthen the evidence for DURA’s computational feasibility within the evaluated prototype ranges.

AuthorsOpenReview ↗

Response to Weakness 2

We appreciate the reviewer’s suggestion to investigate prototype merging. Our additional experiments show that merging reduces prototype counts but decreases All and New ACC on both datasets under both evaluation protocols. Since clustering ACC is the primary evaluation metric in OCD, these results support retaining DURA’s original online discovery and refinement procedure over the evaluated merging strategy. Specifically, using DURA-DINOv2, merging reduces final prototype counts by 36.14% on AquaOV255 and 45.06% on AquaOCD343, while lowering Strict–Hungarian New ACC by 3.30 and 6.44 percentage points, respectively.

Merging strategy. Only newly discovered prototypes can merge; all initial known-category prototypes are protected. We set tau_merge = tau − 0.05, where tau is the original discovery threshold. Every 16 query batches, we merge the pair of new prototypes with the highest cosine similarity above this threshold, update the affected similarities, and repeat until no qualifying pair remains.

The merged prototype is calculated as:

p_merge = normalize(n_a p_a + n_b p_b),

where n_a and n_b are cumulative query-assignment counts. We sum these counts, retain the earlier stable prototype ID, and record an alias for the removed ID. The threshold offset and merging interval were fixed before the experiment and shared across both datasets. Neither query ground-truth labels nor target category counts are used to determine merges.

Table Response2a. Discovery accuracy with and without prototype merging using DURA-DINOv2. ACC values are percentages, reported as mean ± sample standard deviation across five query permutations.

DatasetVariantProtocolAll ACCOld ACCNew ACC
AquaOV255Original DURAGreedy75.70 ± 1.3896.64 ± 0.0763.5 ± 2.15
AquaOV255With mergingGreedy74.07 ± 1.7796.71 ± 0.0460.94 ± 2.78
AquaOV255Original DURAStrict71.42 ± 1.2989.19 ± 0.7161.11 ± 2.01
AquaOV255With mergingStrict69.32 ± 1.7389.14 ± 1.2057.81 ± 2.65
AquaOCD343Original DURAGreedy79.41 ± 0.2395.31 ± 0.0355.54 ± 0.60
AquaOCD343With mergingGreedy78.42 ± 0.3895.80 ± 0.0252.32 ± 0.94
AquaOCD343Original DURAStrict70.08 ± 0.3385.23 ± 0.0747.35 ± 0.79
AquaOCD343With mergingStrict68.88 ± 0.4687.50 ± 3.8640.91 ± 5.82

Table Response2b. Final prototype counts with and without merging. Results are mean ± sample standard deviation; reductions are calculated from the mean counts.

DatasetOriginal DURAWith mergingReduction
AquaOV255768.60 ± 18.02490.80 ± 16.6836.14%
AquaOCD343710.20 ± 4.38390.20 ± 5.8145.06%

These comparisons clarify the trade-off between reducing prototype counts and preserving category discrimination. On AquaOCD343, merging also increases the standard deviation of Strict–Hungarian New ACC from 0.79 to 5.82 percentage points, indicating greater variability across query orders under the evaluated merging strategy.

DURA addresses fragmentation through offline unknown separation (Section 3.3, Eq. (5)), online category alignment, and confidence- and support-weighted prototype refinement (Section 3.4, Eqs. (8) and (12)). These mechanisms are designed to improve representation consistency and limit unreliable prototype drift, helping visually diverse samples from the same category match existing prototypes.

AuthorsOpenReview ↗

Response to Weakness 3

We sincerely appreciate the reviewer’s valuable suggestion. Our additional experiments show that online updates improve alignment for correctly assigned queries, supported prototypes become more class-consistent, and early feedback errors reduce All/New ACC by at most 0.02/0.03 percentage points under the tested settings. Tables Response3a–3c use DURA-DINOv2. Results are mean ± sample standard deviation over five query permutations; ground truth is used only for offline auditing.

Assignment errors. We examine queries matched to existing prototypes. Initial prototypes retain known-class identities; created prototypes receive their creating query’s true class for auditing. An assignment is wrong when query and prototype identities differ. These identities stay fixed for this audit; Table Response3c separately tracks changing dominant classes. These rates differ from Hungarian ACC.

Table Response3a. Wrong-class assignment rates among matched queries (%).

DatasetAll matchedKnown-class matchedNovel-class matched
AquaOV25512.15 ± 0.431.03 ± 0.0418.91 ± 0.71
AquaOCD3438.93 ± 0.361.37 ± 0.0220.51 ± 0.93

The All rate is sample-weighted within each run. Known-class error rates are low; novel-class queries are more challenging.

Update effects. Before updating, we select the most similar available true-class prototype and keep its ID fixed for comparison after encoder adaptation and prototype refinement. Queries without a true-class prototype are excluded.

Table Response3b. Query–true-class-prototype similarity before/after updates.

DatasetAssignmentBeforeAfter
AquaOV255Correct0.72353 ± 0.002260.72414 ± 0.00226
AquaOV255Wrong0.41569 ± 0.007680.41571 ± 0.00769
AquaOCD343Correct0.78789 ± 0.000970.78803 ± 0.00097
AquaOCD343Wrong0.32420 ± 0.006370.32420 ± 0.00638

Correctly assigned queries become more aligned with their true-class prototypes, while incorrectly assigned queries show little change in true-class similarity; their similarity to the assigned wrong prototype increases by only 0.000510 ± 0.000028 and 0.000242 ± 0.000014, respectively, indicating small immediate effects whose downstream consequences we examine below.

Prototype consistency. For created prototypes receiving ≥20 assignments, consistency is the dominant true-class proportion within the first or last 10 assignments. The dominant class may change between windows.

Table Response3c. Prototype consistency over time.

DatasetPrototypes/runFirst 10 (%)Last 10 (%)Dominant class changed and consistency improved (%)
AquaOV255212.4 ± 12.757.82 ± 1.3582.16 ± 1.0831.74 ± 2.12
AquaOCD343178.6 ± 8.353.67 ± 1.0979.03 ± 1.2427.89 ± 1.85

Consistency rises by 24.34/25.36 percentage points, including prototypes that become more consistent while changing their dominant class.

Early-error perturbations. Using DURA-DINO (DINOv1), we randomly replace feedback IDs with different existing prototype IDs for 5%, 10%, or 20% of queries in the first 10% of each stream. Only adaptation/refinement IDs are changed; recorded predictions remain intact, while subsequent predictions use the updated model and memory. Five paired permutations share checkpoints, initial states, and query orders with the 0% baseline. ACC covers the unperturbed last 90%; final counts cover the full stream.

Table Response3d. Early-feedback perturbation results. ACC (%) is shown as Greedy / Strict; prototype counts are shared across protocols.

DatasetRateAll ACC (G/S)Old ACC (G/S)New ACC (G/S)Final prototypes
AquaOV2550%60.90±0.67 / 56.20±0.6893.11±0.19 / 87.01±1.3342.15±1.03 / 38.27±0.511230.20±8.93
AquaOV2555%60.89±0.68 / 56.19±0.6993.12±0.18 / 87.02±1.3442.13±1.04 / 38.25±0.511230.40±8.96
AquaOV25510%60.88±0.65 / 56.18±0.6693.11±0.19 / 87.01±1.3342.12±0.99 / 38.24±0.471230.20±9.07
AquaOV25520%60.96±0.59 / 56.26±0.6593.11±0.19 / 87.01±1.3342.25±0.92 / 38.36±0.331230.80±9.31
AquaOCD3430%75.03±0.32 / 65.12±0.2694.64±0.05 / 84.77±0.1645.55±0.80 / 35.58±0.79705.20±7.26
AquaOCD3435%75.03±0.32 / 65.12±0.2694.64±0.05 / 84.77±0.1645.55±0.80 / 35.57±0.79705.20±6.87
AquaOCD34310%75.07±0.35 / 65.16±0.3094.64±0.05 / 84.77±0.1645.65±0.91 / 35.68±0.91704.80±7.50
AquaOCD34320%75.07±0.35 / 65.16±0.2994.64±0.04 / 84.77±0.1545.65±0.89 / 35.67±0.90704.80±7.50

Maximum mean All/New ACC decreases are 0.02/0.03 points on AquaOV255 and 0.00/0.01 points on AquaOCD343; mean final counts change by at most 0.60/0.40, respectively.

Reviewer_v5vdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies underwater on-the-fly category discovery, in which a model trains offline using only a labeled support set of known categories and subsequently makes predictions sample by sample on an unlabeled stream containing both known and novel categories. DURA combines two stages. During offline training, it uses an angular-margin prototype classification loss and generates pseudo-unknown samples through Mixup and MKEE adopted from LTC. A maximum-similarity penalty separates these samples from known-category prototypes. During online inference, it expands the prototype memory using a similarity threshold. After completing predictions for the current batch, it updates the affine parameters of the normalization layers in the final encoder block and then refines the prototypes according to assignment confidence and sample counts. The paper explicitly states that updates affect only subsequent queries and do not retrospectively revise recorded predictions (pp. 3–7, 14–15).

The paper also introduces AquaOCD343, containing 343 categories and 67,651 images, and combines it with four existing underwater datasets to form UOCDBench. Experiments cover five underwater datasets and five standard datasets, reporting All, Old, and New accuracy under two matching protocols. The evaluation also includes component ablations, a dual-stage ablation, category-count estimation, and hyperparameter analyses. The reported results broadly support the method’s potential, although some aspects of metric calculation and the stability of the gains require clarification.

Critical strengths
  1. The underwater benchmark has potential research value, and the evaluation includes a large new dataset. UOCDBench brings together five datasets under a unified protocol, covering different category scales and levels of fine-grained distinction. AquaOCD343 contains 343 categories and 67,651 images and explicitly documents its long-tailed distribution, with 2–7,760 images per category. Even if the method is subsequently surpassed, the dataset and evaluation protocol retain independent value for underwater open-world recognition.

  2. The improvements in novel-category accuracy over the strongest baseline, TALON, are convincing under the same backbone, and the ablations support the complementary roles of the two stages. In the DINOv2 comparison in Table 13, novel-category accuracy on CUB-200 increases from 37.2% to 62.9%. Table 8 shows that offline unknown supervision alone increases Strict–Hungarian novel-category accuracy from 36.7% to 46.3%, substantiating the motivation to prepare representations for unknown categories during offline training.

Critical weaknesses
  1. The paper reports extensive experimental results, but several details affecting their interpretation and reproducibility remain unclear. Equation (13) provides a generic clustering-accuracy formula without sufficiently specifying the matching and aggregation procedures for All, Old, and New under the two evaluation protocols. Given the support/query partitions in Table 5, and assuming consistent sample-level scoring and matching across columns, some ALL scores in Table 1 cannot be recovered by weighting the corresponding Old and New scores. The authors therefore need to explain the exact evaluation procedures. Table 13 already provides same-backbone comparisons on standard datasets, but the corresponding configurations for the underwater experiments still require clarification.

  2. The complete dual-stage ablation in Table 8 is valuable and supports the contribution of offline unknown supervision. However, adding online adaptation to offline unknown supervision improves All accuracy by 0.8 and 0.6 percentage points and new accuracy by 1.6 and 1.1 percentage points under Greedy–Hungarian and Strict–Hungarian, respectively. The paper does not report the variability of these differences across training seeds or query orders, making it difficult to assess whether the additional gains are stable. Furthermore, Table 12 shows that changing the novel-prototype update rate from 0.10 to 0.05 or 0.15 reduces Strict–Hungarian New accuracy from 47.4% to 34.2% or 33.6%. Although the paper provides parameter-sensitivity experiments, it does not clearly explain the validation protocol used for parameter selection.

Questions for authors
  1. Please clarify the relationship between the evaluation metrics and the numerical results reported in the tables.

  2. Please specify the configurations used for the underwater comparisons and online inference.

  3. Please explain the basis for parameter selection and provide evidence demonstrating the stability of the online adaptation gains.

4 comments
AuthorsOpenReview ↗

Response to Question 1

We sincerely appreciate the reviewer’s thoughtful feedback and apologize for our late response. We took additional time to complete the training seed experiments and carefully verify the reproducibility of all reported results.

Response to Question 1

We sincerely appreciate the reviewer’s careful examination of the evaluation metrics and reported results. We apologize for the reporting errors and insufficient explanation in the original manuscript. We have expanded Appendix A.8, “Detailed Evaluation Metrics,” to clarify how the matching procedures produce the All/Old/New values in Tables 1.

Relationship between the metrics and table entries. Each All/Old/New triplet is calculated from the same query predictions and checkpoint. Old and New queries are separated according to their ground-truth category membership, not their predicted prototype IDs. Accuracy measures the proportion of correctly matched samples after aligning predicted prototype IDs with ground-truth category labels.

The two protocols differ in how this alignment is obtained:

  • Greedy–Hungarian: Hungarian matching is optimized independently within the Old and New query subsets. Old ACC and New ACC use their respective optimal mappings, which may assign the same prototype ID to different categories. All ACC combines the correctly matched samples from these two evaluations; it is not calculated using another global matching.
  • Strict–Hungarian: Hungarian matching is optimized once over the complete query set. This single mapping is then held fixed when calculating Old and New ACC. All ACC is the correctly matched fraction over the complete query set under that same mapping.

Under either protocol, let C_old and C_new denote the numbers of correctly matched Old and New queries, and let N_old and N_new denote their total sample counts. Then:

ACC_Old = C_old / N_old

ACC_New = C_new / N_new

ACC_All = (C_old + C_new) / (N_old + N_new)

Equivalently:

ACC_All = (N_old × ACC_Old + N_new × ACC_New) / (N_old + N_new)

Thus, ACC_All is a weighted aggregation of ACC_Old and ACC_New, rather than their unweighted average. The query counts determine the weights; the matching protocol determines which samples count as correct. For a fixed dataset partition, the weights are identical across methods and protocols, even though the resulting accuracies differ.

Table 14 in Appendix A.8 explicitly provides the Old/New query counts: AQUA20 2,238/2,084; AquaOV255 5,566/9,592; MAS3K 797/1,316; USIS16K 5,344/5,463; and AquaOCD343 25,376/16,899. These counts, rather than support-set sizes or category proportions, determine the aggregation of every corresponding triplet in Tables 1. All metrics are sample-level accuracies, not class- or batch-averaged scores. Calculations use unrounded results, and percentages are rounded only for presentation, so recomputing All from the displayed Old/New values may produce small rounding differences.

Corrections and verification. During manual preparation of the tables, outdated query-subset proportions from preliminary MAS3K and USIS16K configurations were inadvertently used in some All ACC calculations. We sincerely apologize for these mistakes.

We have corrected the affected results and rechecked the numerical entries, aggregation calculations, and best/second-best markings against the final reported configurations.

Reproducibility and release of research materials. We have provided the complete training and evaluation code, dataset-split files, hyperparameter settings, random seeds, and experimental configurations in an anonymous repository(https://anonymous.4open.science/r/anonymous-code114514/). These materials specify the settings needed to reproduce the results on each dataset, including the evaluation procedures in Appendix A.8 and the same-backbone comparisons in Appendix A.9. We have verified through reproduction runs that the released code and configurations reproduce the reported results across the evaluated datasets.

We have not yet identified a suitable way to distribute the model weights anonymously; they will be released after the review process concludes. AquaOCD343 includes privately collected data and cannot be publicly released before publication. Its complete split definitions and experimental settings are nevertheless included in the repository to document the evaluation and support reproduction once the dataset becomes available.

AuthorsOpenReview ↗

Response to Question 2

We thank the reviewer for pointing out that the backbone configurations and online inference settings for the underwater comparisons required more explicit documentation. We have added Appendix A.9, “Comparison under Different Pretrained Backbones on UOCDBench,” and Table 15, complementing the same-backbone comparisons on standard datasets in Appendix A.7 and Table 13.

Because previous methods do not report results on UOCDBench, we evaluate them using their publicly available implementations and published configurations. All methods use the same UOCDBench support/query partitions and the same Greedy–Hungarian and Strict–Hungarian protocols.

In the main Table 1, SLC, RankStat, SMILE, PHE, and LTC use CLIP ViT-B/16, while TALON uses DINOv2 ViT-B/14. DURA is evaluated with CLIP ViT-B/16, DINOv1 ViT-B/16, and DINOv2 ViT-B/14. Appendix A.9 additionally provides controlled same-backbone comparisons.

With the backbone fixed to DINOv2, DURA achieves higher All and New ACC than TALON on all five underwater datasets under both protocols. For example, on AquaOCD343, DURA improves All ACC by 4.4/4.5 percentage points and New ACC by 9.7/10.5 points under Greedy/Strict–Hungarian, respectively. This confirms that the improvements are not attributable only to the choice of pretrained backbone.

DURA offline configuration. All DURA variants are trained for 100 epochs with a batch size of 16 using AdamW. The weight decay is 1 × 10⁻⁴, and the learning rates for the trainable backbone parameters and category prototypes are 1 × 10⁻⁵*and 5 × 10⁻⁴, respectively. We use cosine learning-rate scheduling with a 10-epoch warm-up initialized at 1 × 10⁻⁶. The scale and angular-margin parameters are 30 and 0.4.

For pseudo-unknown generation, we use α_mix = 0.4, ε = 0.05, and λ_ρ = 0.1. One pseudo-unknown representation is generated for each labeled sample. The density references are extracted from the current labeled batch, and the Gaussian RBF bandwidth is determined by the median pairwise distance between generated and reference representations.

Online inference settings. As detailed in Section 3.4 and Appendix A.2, known-category prototypes are initialized from normalized support-set class means, and queries are processed sequentially within each batch. A query is assigned to its most similar prototype when the similarity reaches the discovery threshold; otherwise, a new prototype is created immediately. Predictions are recorded before updating and are not revised retrospectively. After each batch, one SGD step updates only the affine parameters of the normalization layers in the final Transformer block, using learning rates of 0.01 , similarity scaling of 30, and gradient clipping at 1.0. Prototype refinement then uses the pre-update query representations, with known/novel update rates of 0.02/0.10 and damping parameters of 32/8; prototypes supported by only one query are not refined. The discovery threshold is the 0.001 quantile of support-sample maximum similarities and is capped at 0.65. Query ground-truth labels are used only for evaluation. All experiments are conducted on a single NVIDIA RTX 5090 GPU.

AuthorsOpenReview ↗

Response to Question 3

We sincerely appreciate the reviewer’s suggestion. The additional query-order and training-seed experiments provide complementary evidence that online adaptation delivers stable, repeatable gains. Full DURA achieves higher mean All and New ACC under both protocols in both experiments. Across five paired training runs, New ACC improves in every pair under both protocols, further supporting the reliability of the online adaptation gains.

How the parameters were selected. We compared candidate values using the sensitivity experiments in Appendix A.6, Tables 9–12. Selection was based on query-set All, Old, and New ACC, with particular attention to novel-category discovery. The resulting settings are: unknown-separation margin 0.2, unknown-loss weight 0.1, discovery-threshold cap 0.65, and novel-prototype update rate 0.10. These values were fixed before the repeated experiments below and were not retuned for individual training seeds or query orders.

Stability across query orders. We evaluated DURA-DINOv2 on AquaOCD343 using five paired random query permutations. Within each pair, Full DURA and w/o Online Adapt share the same offline checkpoint, initial state, and query order. Both retain offline unknown supervision and normal prototype creation; the control disables online representation adaptation and prototype refinement.

Table Response5. Online-adaptation gains across five paired query permutations. ACC values are percentages, reported as mean ± sample standard deviation.

ProtocolSettingAll ACCOld ACCNew ACC
Greedy–HungarianFull DURA79.43 ± 0.2295.33 ± 0.0355.55 ± 0.60
Greedy–Hungarianw/o Online Adapt78.63 ± 0.2095.11 ± 0.0253.88 ± 0.53
Strict–HungarianFull DURA70.11 ± 0.3185.24 ± 0.0747.39 ± 0.88
Strict–Hungarianw/o Online Adapt69.48 ± 0.2484.93 ± 0.1246.28 ± 0.78

Full DURA achieves higher mean accuracy in all six metrics. All ACC increases by 0.80/0.63 percentage points, and New ACC increases by 1.67/1.11 points under Greedy/Strict–Hungarian, consistent with the improvements reported in Table 8.

Stability across training seeds. We additionally completed five paired training runs using DURA-DINOv2 on AquaOCD343*, with seeds 1028–1032, a fixed dataset partition, training batch size 128, and query batch size 64. Within each pair, both variants share the pretrained initialization, random head initialization, initial prototypes, training order, and query order. Each seed uses its corresponding query permutation, shared by the two variants.

Both variants are trained for 100 epochs without prototype merging. Each independently selects the checkpoint with the highest query-set Greedy New ACC; all six metrics come from that checkpoint’s prediction sequence. These runs follow the complete training-and-evaluation procedure, including retaining Full DURA’s online encoder updates for subsequent training epochs.

Table Response6. Stability of online-adaptation gains across five training seeds on AquaOCD343 using DURA-DINOv2. Results are mean ± sample standard deviation. ACC values are percentages; paired gains Δ = Full DURA − w/o Online Adapt are measured in percentage points.

ProtocolSettingAll ACCOld ACCNew ACC
Greedy–HungarianFull DURA79.776 ± 0.22595.674 ± 0.22355.904 ± 0.379
Greedy–Hungarianw/o Online Adapt79.441 ± 0.34895.797 ± 0.13854.880 ± 0.694
Greedy–HungarianPaired gain Δ+0.335 ± 0.205−0.123 ± 0.269+1.024 ± 0.391
Strict–HungarianFull DURA71.425 ± 0.84085.382 ± 0.83650.467 ± 1.092
Strict–Hungarianw/o Online Adapt70.531 ± 1.04085.374 ± 0.62548.241 ± 2.033
Strict–HungarianPaired gain Δ+0.895 ± 1.421+0.008 ± 1.335+2.226 ± 1.894

Paired gains and their standard deviations are calculated directly from the five within-pair differences. New ACC improves in 5/5 pairs under both protocols, with mean gains of 1.024/2.226 percentage points. All ACC improves in 5/5 Greedy pairs and 4/5 Strict pairs, providing additional evidence that the observed benefits recur across repeated training runs.

Reviewer_v5vdOpenReview ↗

I will keep my rating.

It did not fundamentally dispel my doubts. I will keep my rating.

Reviewer_52XfOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DURA, a dual-stage framework for underwater On-the-Fly Category Discovery (OCD). It combines offline pseudo-unknown separation with online representation adaptation and prototype refinement to improve novel category discovery. The authors also introduce AquaOCD343 and UOCDBench for evaluating underwater OCD. Experiments on underwater and standard benchmarks demonstrate competitive performance.

Critical strengths
  1. The paper evaluates DURA on multiple underwater and standard OCD benchmarks, demonstrating competitive performance, particularly for novel-category accuracy.

  2. The introduction of AquaOCD343 and UOCDBench provides additional evaluation resources for underwater category discovery.

Critical weaknesses
  1. The framework largely combines existing techniques. Pseudo-unknown generation follows LTC, while online representation adaptation builds upon TALON. The additional unknown-separation loss and confidence-based prototype updates appear relatively incremental.

  2. The contributions of several components are modest. In Table 3, removing online adaptation reduces overall accuracy by only 0.8 and 0.6 percentage points under Greedy–Hungarian and Strict–Hungarian, respectively. Removing representation adaptation produces even smaller differences.

  3. The experiments do not sufficiently demonstrate why the proposed combination is preferable to simpler alternatives or whether its additional complexity is justified.

  4. Despite the reported clustering accuracy, DURA still substantially misestimates the number of discovered categories. In Appendix Table 7, DURA-DINOv2 predicts 710 categories on AquaOCD343, compared with only 343 ground-truth categories. Similar overestimation occurs on AquaOV255.

  5. These results raise concerns about excessive category fragmentation, which directly affects the central objective of on-the-fly category discovery.

  6. The paper also lacks systematic evaluation of robustness to different query orders and accumulated pseudo-labeling errors, despite their importance for sequential category discovery.

Questions for authors
  1. What are the fundamental methodological differences between DURA and the combination of LTC-style unknown generation and TALON-style online adaptation? Can the authors further justify the necessity of the additional components?

  2. Why does DURA still substantially overestimate the number of categories on AquaOCD343 and AquaOV255? How does the method address category fragmentation during online discovery?

  3. How sensitive is DURA to different query arrival orders and incorrect early category assignments? Can the authors provide quantitative robustness results across multiple random query streams?

4 comments
AuthorsOpenReview ↗

Response to Weaknesses 4 and 5 and Question 2

We appreciate the reviewer’s concern about category-count overestimation and fragmentation. DURA substantially reduces fragmentation relative to the compared methods, and our additional merging experiment shows that further reducing prototype counts can sacrifice discovery accuracy. Table 7 shows that DURA variants provide the closest category-count estimates across all five datasets. Explicit merging reduces prototype counts by 36.14% on AquaOV255 and 45.06% on AquaOCD343, but lowers Strict–Hungarian New ACC by 3.30 and 6.44 percentage points, respectively. Since clustering ACC is the primary evaluation metric in OCD, these results support retaining the original DURA procedure, which achieves higher All and New ACC on both datasets.

Why overestimation remains. In OCD, novel categories are absent from offline training, their number is unknown, and queries arrive sequentially without labels. Exact category-count recovery is therefore challenging. AquaOCD343 and AquaOV255 further contain many categories, long-tailed distributions, and substantial within-class appearance variation, making visually different samples from the same category more likely to trigger separate prototype creation.

How DURA reduces fragmentation. DURA combines complementary offline and online designs. Offline unknown separation prepares representations for distinguishing unfamiliar samples from known categories. During online discovery, representation adaptation improves alignment with the emerging category structure, while confidence- and support-aware prototype refinement limits unreliable updates and strengthens representations supported by recurring observations. These mechanisms help subsequent samples match existing prototypes rather than repeatedly creating additional ones.

Our additional direct LTC+TALON comparison provides further evidence of this improvement. Using the same DINOv1 backbone and query order, DURA reduces the final prototype count from 4,415 to 703 on AquaOCD343 and from 2,653 to 1,226 on AquaOV255, while improving discovery accuracy.

Explicit prototype-merging experiment. We additionally evaluated merging with DURA-DINOv2. Only newly discovered prototypes can merge; all initial known-category prototypes are protected. The merging threshold is fixed at tau_merge = tau − 0.05, and merging is performed every 16 query batches, without using query labels or a target category count.

Table Response7. Effect of prototype merging on category counts and Strict–Hungarian accuracy. ACC values are percentages; results are mean ± sample standard deviation across five query permutations.

DatasetVariantFinal prototypesAll ACCOld ACCNew ACC
AquaOV255Original DURA768.60 ± 18.0271.42 ± 1.2989.19 ± 0.7161.11 ± 2.01
AquaOV255With merging490.80 ± 16.6869.32 ± 1.7389.14 ± 1.2057.81 ± 2.65
AquaOCD343Original DURA710.20 ± 4.3870.08 ± 0.3385.23 ± 0.0747.35 ± 0.79
AquaOCD343With merging390.20 ± 5.8168.88 ± 0.4687.50 ± 3.8640.91 ± 5.82

The comparison demonstrates the trade-off between fewer prototypes and preserving category discrimination. On AquaOCD343, merging also increases the standard deviation of Strict–Hungarian New ACC from 0.79 to 5.82 percentage points, indicating greater variability across query orders under the evaluated merging strategy.

The rebuttal has addressed several of my original concerns and strengthened the empirical evidence for DURA. While the remaining category-fragmentation issue limits my confidence, I acknowledge the additional experimental support for the proposed framework and am inclined to update my rating from 2 to 3.

AuthorsOpenReview ↗

Our Sincere Thanks for Your Time, Thoughtful Feedback, and Careful Reassessment

We sincerely appreciate your careful reassessment and recognition of the additional experimental evidence. Your constructive feedback helped us strengthen the experiments and clarify DURA’s contributions. Thank you again for your time, thoughtful feedback, and willingness to reconsider your assessment.

Thank you for the authors’ response.

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

Related papers

Discussion (0)

Sign in to comment.