acceptodds
Under review as a conference paper at ICLR 2027

Active-SAOOD: Active Sparsely Annotated Oriented Object Detection in Remote Sensing Images

Abstract

Sparse annotation has emerged as an effective strategy for reducing the annotation burden of oriented object detection in remote sensing images. However, with a limited annotation budget, selecting informative instances becomes critical, while random selection may allocate annotations to less valuable instances. To this end, we propose an active learning-based sparsely annotated oriented object detection (SAOOD) method, termed Active-SAOOD, which actively selects high-value instances for annotation. Specifically, existing image-level active learning methods face two limitations in SAOOD task: (1) the granularity mismatch in annotation value evaluation and (2) unstable training caused by excessive selection of difficult samples without considering model states. To overcome these challenges, we design instance-level evaluation dimensions tailored to SAOOD, enabling the selection of instances that improve both fundamental detection and representation generalization abilities. Moreover, a model-state observation strategy is introduced to identify instances that better match the current learning state of model, thereby improving training stability. Extensive experiments on multiple datasets demonstrate that the Active-SAOOD consistently improves detection performance and training stability under different sparse annotation ratios, highlighting its practical value for SAOOD. The code will be public.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Oriented object detection, Sparsely annotation, Remote sensing images

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_843BOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Active-SAOOD, an instance-level active learning framework for sparsely annotated oriented object detection. It evaluates annotation value using detection uncertainty and feature diversity, and introduces a model state observation strategy to select suitable instances during training.

Critical strengths
  1. The motivation is clear, and instance-level active learning is well suited to the sparse annotation setting.
  2. The method achieves promising improvements across different datasets and annotation ratios.
  3. The ablation studies and repeated experiments provide good evidence of its effectiveness and training stability.
Critical weaknesses
  1. While individual components are evaluated, a comparison between detection uncertainty, representation diversity, and their combination would better demonstrate their complementary benefits.
  2. The linear weighting in Eq. (11) seems somewhat heuristic. An ablation with alternative weighting strategies would help justify this design.
  3. The method relies on predicted candidate instances, especially under sparse annotations. An experiment on how candidate quality affects detection performance would be helpful.
Questions for authors

Please refer to the concerns raised in the Weaknesses section.

2 comments

Q1: Ablation study on detection uncertainty and representation diversity
Thank you for your suggestions. We have conducted additional ablation experiments on detection uncertainty and representation diversity. As shown in the table, removing either detection uncertainty or representation diversity leads to a substantial performance drop. Combining both achieves the best performance of 35.79% mAP. These results demonstrate that the detection ability and representation generalization are complementary, and jointly considering both leads to better performance.

Detection uncertaintyRepresentation diversitymAP (%)
✓31.53
✓32.21
✓✓35.79

Q2: Experiment on alternative weighting strategy
We compare alternative weighting strategies, as shown in the table. Linear and nonlinear weighting achieve close performance. In fact, Eq. (11) is simply a straightforward choice, as other formulations that avoid introducing too many difficult-to-learn instances when the model is weak could also satisfy this requirement.

Weighting StrategySelection FormulamAP (%)
Nonlinear weightingexp(−(S−Aˉ)2)exp\left(-(S-\bar A)^2\right)35.68
Linear weightingAˉS+(1−Aˉ)(1−S)\bar{A}S+(1-\bar{A})(1-S)35.79

More importantly, the key component of Eq. (11) is the ability-aware instance value SS, defined in Eq. (10), which dynamically weights instance values according to model's current abilities across dimensions, thereby selecting instances that better improve the model's abilities in the corresponding dimensions. This enables instance valuation to consider not only instance information but also model's abilities and learning needs. To further verify this, we replace ability-aware instance value in Eq. (10) with a direct summation of active learning metrics (common practice, i.e., ∑Ui\sum U_i). This results in a substantial performance drop of 2.97%, demonstrating the importance of ability-aware instance value.

Q3: Effect of candidate quality on detection performance
Thank you for your suggestion. We use three models with different performance to generate candidate instances of varying quality and further evaluate the quality of these candidates. We then apply our Active-SAOOD to select the same number of instances from each candidate pool to ensure a fair comparison. The selected instances are combined with the initial 1% labeled data to train the S2Teacher, and the final mAP is reported. As shown in the table, higher candidate quality leads to greater performance gains from active learning. However, the differences in gains across different candidate quality levels are relatively small, and even low-quality candidates can bring significant improvements. This suggests that when the model's ability is limited, selecting relatively valuable instances can already yield substantial performance gains. As the model becomes stronger, it can generate higher-quality candidates, enabling active learning to select more valuable instances and further improve the overall performance. Thus, candidate quality and the model ability can mutually reinforce each other.

Candidate QualityCandidate mAP on training set (%)Candidate Recall (%)Final mAP (%)
Low42.074.646.50
Medium68.592.346.65
High74.495.546.83
Reviewer_843BOpenReview ↗

Response to the authors

Thank you for your response, which has fully addressed my concerns.

Reviewer_f7XWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Active-SAOOD, an instance-level active learning framework for sparsely annotated oriented object detection in remote sensing images. Rather than randomly annotating objects or selecting entire images, the method identifies informative individual instances under a limited annotation budget. The proposed sampling strategy combines classification uncertainty, predicted localization and orientation errors, inter-class diversity, and intra-class feature diversity. A Model State Observation (MSO) mechanism dynamically weights these criteria according to the detector's current capabilities and aims to progressively shift sample selection from easier to harder instances. The framework integrates these components into a teacher-student detection pipeline with pseudo-label learning. Experiments on DOTA-v1.0, DOTA-v1.5, DIOR, and SODA-A demonstrate improvements over random annotation and several active learning baselines. For example, with 1% annotations on DOTA-v1.0, the S2 Teacher-based implementation improves from 30.73% to 35.79% mAP. Additional ablations and repeated experiments support the contributions of individual sampling criteria and improved training stability.

Critical strengths

1.Meaningful improvements under limited annotation budgets. The results in Table 1 demonstrate substantial performance improvements when annotations are extremely limited. At 1% annotation on DOTA-v1.0, Active-SAOOD improves S2 Teacher from 30.73% to 35.79% mAP and RSST from 23.19% to 38.54% mAP. These improvements address an important practical problem in annotation-efficient object detection.

2.Task-aware instance selection with supporting ablations. The method explicitly accounts for classification, localization, orientation, and feature diversity instead of relying on a single uncertainty score. The orientation-sensitive component is particularly relevant for rotated bounding boxes. Table 3 and Appendix Table 6 demonstrate that removing individual components degrades detection performance or the corresponding detection capabilities.

3.Evidence of improved stability and generalization. The evaluation includes multiple datasets, annotation budgets, and repeated experiments. Table 4 reports substantially lower performance variance for Active-SAOOD than for random sparse annotation. The supplementary experiments also examine MSO using an independent held-out set, strengthening the evidence that the observed improvements are not restricted to a particular model-state evaluation setting.

Critical weaknesses

Limited methodological novelty and missing comparisons with closely related work. The proposed framework primarily combines established active learning concepts, including uncertainty-based sampling, diversity-based selection, class balancing, and adaptive sample weighting. More importantly, instance-level active learning and predicted-error-based selection have already been explored in prior work: Yoo and Kweon, Learning Loss for Active Learning (CVPR 2019), introduced a learned loss-prediction module for selecting informative unlabeled samples. Lyu et al., Box-Level Active Detection (CVPR 2023), introduced ComPAS, which selects individual objects for annotation and combines human labels with pseudo-labels. Box-Level Class-Balanced Sampling for Active Object Detection (ICIP 2024) already explored class balancing in box-level active learning. The related work and experimental comparisons do not sufficiently establish how Active-SAOOD fundamentally advances beyond these approaches. In particular, the claim that existing active learning methods primarily operate at the image level overlooks relevant box-level frameworks. Comparisons against image-level methods alone are insufficient to establish the novelty of instance-level selection. Potential mathematical weakness in the proposed MSO sampling mechanism. Equation (11) defines the final sampling score as: S_final = A_bar * S + (1 - A_bar) * (1 - S). This can be rewritten as: S_final = (1 - A_bar) + (2 * A_bar - 1) * S. Consequently, when A_bar = 0.5, every instance receives the same final score, regardless of its uncertainty or diversity. When A_bar is below 0.5, the ranking favors lower values of S, while above 0.5 the ranking reverses. Thus, the formulation produces a ranking reversal at a fixed threshold rather than a genuinely smooth transition from easy to difficult samples. This observation challenges the claimed continuous adaptation to the model's learning state. The authors should provide additional theoretical justification and demonstrate that the proposed formulation is preferable to simpler or smoother curriculum-based alternatives. Insufficient validation of instance uncertainty estimation under extremely sparse supervision. The localization and orientation uncertainty prediction (LUP) module is trained to estimate regression errors using the limited labeled instances. However, the paper does not directly demonstrate how accurately these predicted uncertainties correspond to actual localization and orientation errors on previously unseen instances. This is important because the central motivation is to improve annotation selection when supervision is extremely limited. A poorly calibrated uncertainty estimator could select uninformative instances, while confidently incorrect predictions and entirely missed objects may never be selected. Although the ablations demonstrate the usefulness of the combined localization-orientation criterion, they do not establish the reliability of the uncertainty predictions themselves. Additional analysis of prediction-error correlation, uncertainty calibration, and candidate proposal recall would substantially strengthen the claims. A secondary experimental concern is that the 1% results in Table 1 coincide with the maximum values reported across five runs in Table 4, rather than the corresponding means. The authors should clarify the reporting protocol and provide mean and standard deviation results consistently across all annotation budgets.

Questions for authors
  • Novelty and comparison with existing box-level active learning: How does Active-SAOOD differ fundamentally from ComPAS (CVPR 2023) and existing class-balanced box-level active learning methods? Can the authors provide comparisons under identical annotation budgets, detector architectures, and pseudo-labeling settings?

  • MSO mathematical formulation: Equation (11) assigns identical scores to all instances when the average model ability equals 0.5 and reverses sample rankings across this threshold. How do the authors reconcile this behavior with the claimed smooth, progressive easy-to-hard sampling mechanism? Would a continuously varying curriculum formulation achieve similar or better results?

  • Reliability of uncertainty prediction: How accurately does the LUP module estimate actual localization and orientation errors when trained with only 1% labeled instances? Can the authors report correlations between predicted uncertainty and true regression error on held-out objects? Additionally, how does the sampling strategy handle objects missed entirely by the teacher detector?

  • Experimental reporting: Do the main results represent individual runs, best-performing runs, or averages across random seeds? Given that Table 1 values coincide with maxima in Table 4, reporting mean and standard deviation across annotation budgets would help establish the statistical reliability of the improvements.

4 comments

Q1: Comparison with box-level active learning
Thank you for the comment. We compare box-level active learning methods under 1% annotation, using S2Teacher for pseudo-label mining across all methods for fairness. Active-SAOOD outperforms the existing methods, demonstrating its effectiveness for remote sensing SAOOD task.

MethodmAP (%)
ComPAS30.78
Class-balanced31.35
Active-SAOOD35.79

These methods are mainly designed for horizontal object detection in natural scenes and do not consider the characteristics of remote sensing SAOOD. Specifically, ComPAS [CVPR23] evaluates instance difficulty via classification and localization disagreements across multiple views and selects the most difficult instances, the class-balanced method [ICIP24] uses category-balanced weights, and Learning Loss [CVPR19] selects samples by loss values. In contrast, Active-SAOOD designs task-specific evaluation dimensions, including orientation uncertainty for oriented object detection, and intra-class diversity to capture the rich feature variations among instances of the same category in remote sensing. Unlike the class-balanced method, which considers only inter-class balancing, our method incorporates both inter-class and intra-class diversity.

More importantly, existing methods typically select instances based solely on sampling metrics, overlooking model's abilities across dimensions. In contrast, our MSO dynamically adjusts sampling preferences across dimensions based on the model's current abilities, prioritizing instances that can better improve weaker ability dimensions. Thus, Active-SAOOD integrates model abilities and learning needs into instance valuation, enabling model ability-aware instance selection.

Q2: The issue with Eq. (11)
Eq. (11) has a critical point only when the model ability Aˉ\bar{A} is exactly 0.5. Any slight deviation (e.g., 0.501) preserves the score distinctions among instances. In our implementation, when Aˉ=0.5\bar{A}=0.5, we select instances using the original ranking based on S. Once newly selected instances are added to training, the model ability is updated and moves across this critical point. Thus, this issue does not affect practical execution. For mathematical completeness, we added an explicit formulation of Eq. (11) at the critical point.

We further compare Eq. (11) with an alternative continuously varying curriculum-like formulation. The close performance indicates that the specific formulation is not critical, as long as it avoids introducing too many difficult instances when model is weak. Eq. (11) is a straightforward choice that favors easy-to-learn instances when the model is weak and challenging ones as it becomes strong.

StrategySelection FormulamAP (%)
Alternativeexp(−(S−Aˉ)2)exp(-(S-\bar A)^2)35.68
OursAˉS+(1−Aˉ)(1−S)\bar AS+(1-\bar A)(1-S)35.79

More importantly, the key component of Eq. (11) is the ability-aware instance value SS, defined in Eq. (10), which assigns higher weights to weaker model ability dimensions, prioritizing instances that can better improve these abilities. Compared with directly summing dimension-wise scores (common practice), our ability-aware valuation improves mAP by 2.97%, showing that SS represents a model-state-aware instance value rather than conventional difficulty. We will revise descriptions for clarity.

Q3: Reliability of LUP prediction and missed objects
We evaluated LUP under 1% annotation, achieving 76.9% ranking accuracy and a 0.783 correlation coefficient, demonstrating that LUP can reliably estimate relative uncertainties under limited supervision. These metrics compare predicted uncertainties with actual regression errors on held-out objects. Since the uncertainties are continuous values rather than discrete labels, we use ranking accuracy to evaluate its ability to distinguish relative uncertainty among instances.

MetricULocθU_{Loc_\theta}
Accuracy (%)76.9
Correlation0.783

Objects missed by the teacher are an inherent limitation of detector-based active learning under sparse supervision, rather than our sampling strategy. Our experiments show close final mAP of 46.50% and 46.83% when the teacher candidate recall is 74.6% and 95.5%, indicating that candidate recall has a limited impact on active learning gains. Moreover, active learning improves teacher detection, enabling recovery of previously missed objects and forming a mutually reinforcing process, pseudo-label mining can also further recover missed objects.

Q4: Why Table 1 reports maximum values
Table 1 reports the max results because baseline without active learning is prone to training failures under 1% annotation, resulting in low min value pulling down the average. Reporting average result may underestimate baseline and exaggerate Active-SAOOD's gain. For a fairer comparison, we report the max results for all methods in Table 1, while Table 4 provides mean and variance across multiple runs. More results will be provided in the final version.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns, particularly the new comparisons with box-level active learning methods and the quantitative validation of LUP uncertainty prediction. These additions strengthen the empirical evaluation.

However, my main concern regarding methodological novelty remains. While the proposed integration is effective, most components build upon established active learning techniques, and the contribution still appears primarily incremental rather than a substantial methodological advance. Equation (11): The clarification addresses the exact critical point, but the abrupt reversal in instance rankings around 0.5 remains conceptually questionable. The similar performance of the alternative formulation also suggests that the specific curriculum mechanism may not be essential.

Nevertheless, I appreciate the authors' constructive responses and the additional evidence presented. I will carefully reread the rebuttal and reconsider my overall assessment, including the possibility of updating my score, in light of these new results.

Thank you for taking the time to reconsider our work.

(1) We would like to further clarify one of our key contributions: Although our evaluation dimensions build upon existing methods and are tailored to remote sensing SAOOD task, how to effectively combine the evaluation scores across dimensions remains a key challenge. Existing methods commonly combine scores using predefined strategies, such as direct summation, without considering the model's learning needs across dimensions. However, instances with the highest summed scores do not necessarily best meet the model's current learning needs across dimensions. In contrast, we propose a strategy to observe model's current abilities and learning needs across dimensions and adaptively adjusts weights to prioritize the instances that can better improve its weaker ability dimensions, thereby achieving model learning need-guided instance valuation. Our experiment show that this strategy improves 2.97% mAP.

The key differences are as follows:

AspectConventional methodsActive-SAOOD
Evaluation dimensionsClassification and localization difficulty, inter-class balanceTask-specific: orientation uncertainty for oriented detection, inter- and intra-class diversity for remote sensing
Model ability awareness✗✓
Instance valuationInstance informationInstance information + dimension-wise model learning needs
Selection objectiveSelect instances with the highest summed scores: ∑iUi\sum_i U_iSelect instances that improve weaker ability dimensions (Eq. 10): ∑ie1−Ai∑je1−AjUi\sum_i\frac{e^{1-A_i}}{\sum_j e^{1-A_j}}U_i
Key distinctionInstance information-driven valuationModel learning need-guided valuation

To illustrate this difference, we provide a simplified example considering only classification and localization/orientation dimensions, while our method actually considers four dimensions. Assume that the current model has strong classification ability but still needs to improve its localization & orientation abilities.

ItemClassificationLocalization & OrientationConventional method ∑iUi\sum_i U_iActive-SAOOD ∑ie1−Ai∑je1−AjUi\sum_i\frac{e^{1-A_i}}{\sum_j e^{1-A_j}}U_i
Model abilityAcls=0.9A_{\mathrm{cls}}=0.9 (Strong)Aloc/orient=0.3A_{\mathrm{loc/orient}}=0.3 (Weak)--
Instance AUcls=0.1U_{\mathrm{cls}}=0.1Uloc/orient=0.8U_{\mathrm{loc/orient}}=0.80.900.55
Instance BUcls=0.9U_{\mathrm{cls}}=0.9Uloc/orient=0.1U_{\mathrm{loc/orient}}=0.11.000.38
Selected instance--BA

Conventional method selects B due to its higher summed scores, whereas Active-SAOOD selects A by assigning greater importance to localization and orientation, where the model remains weaker and requires learning.

(2) The close performance of alternative formulations suggests that the specific mathematical formulation is flexible, as long as it prevents excessive difficult-to-learn instances from compromising stability, rather than indicating that the curriculum mechanism itself is unnecessary. In fact, MSO incorporates model abilities awareness at two levels. At the dimension level, Eq. (10) assigns greater weights to the weaker ability dimensions and prioritizes instances that can better improve them. At the overall level, Eq. (11) adjusts the preference for instances to avoid introducing excessive difficult-to-learn instances when the model is weak, thereby ensuring stability.

Dear reviewer f7XW, this is a gentle reminder regarding our rebuttal. We would be very grateful if you could take a moment to review our newly responses when you have time. Thank you again for your time and effort.

Reviewer_wa9zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Active-SAOOD, an instance-level active learning framework for detecting rotated targets in sparse-labeled remote sensing images. The method designs five evaluation dimensions and proposes the Model State Observation strategy. The paper conducts experiments on multiple datasets such as DOTA-v1.0, and applies this selection strategy to various SAOOD methods. The results show that actively selecting instances can improve detection performance under different labeling ratios.

Critical strengths
  1. The paper studied a practical annotation cost issue, which has certain application value and practical significance. Moreover, for target-dense remote sensing images, selecting high-value targets based on instances is quite reasonable.
  2. The experimental analysis in the paper is relatively comprehensive. The author tested multiple annotation ratios, datasets, and the basic methods of SAAOD, and provided detailed ablation studies, cost calculation analyses, and visualization results.
Critical weaknesses
  1. Formula 11 will deteriorate when the average model ability approaches 0.5, resulting in no relation with S. Or, when A_bar approaches 0.5, the selection result will be completely opposite. For example, when A_bar = 0.49, S_final(0.1) = 0.508, S_final(0.9) = 0.492. The algorithm selects the simple example; while when A_bar = 0.51, S_final(0.1) = 0.492, S_final(0.9) = 0.501. The algorithm selects the difficult example. A change of 0.02 in the model ability can completely reverse the entire Top-K ranking. The ability values in Table 8 do indeed pass near 0.5, and this degradation does have an impact.
  2. The comparison of the methods seems insufficient. For instance, the ALS Teacher published in April 2026 has used difficulty, information, diversity and direction-related indicators for active label selection in the semi-supervised detection of rotating targets in remote sensing. The authors did not adequately discuss or compare these works, so their innovativeness may be questioned.
  3. The performance improvements highlighted in the paper do not represent the contribution of the active selection itself. For instance, the +14.85 improvement of Active-SAOOD is relative to the normal Rotated FCOS calculation, where the benefits brought by the mixed S2 Teacher pseudo-label training are included. Compared to the actual corresponding S2 Teacher, the active selection only increased from 30.73 to 35.79, which is a 5.06 points increase.
  4. Some minor matters. I would like to know why the author chose DOTA 1.0 and 1.5 instead of DOTA 2.0. Furthermore, it is suggested that the author refine the presentation of the example images in Figures 1 and 2. For instance, the several images in the "Labeled Instances" section of Figure 2 appear rather disorganized. The data used in the paper should be the rotated box data. Therefore, it is recommended to change "DIOR" to "DIOR-R". The 1% data in Table 1 is entirely based on the max values from Table 4.
Questions for authors

I have no further questions. I hope the author can provide their responses to the "Critical Weaknesses".

4 comments

Q1: The critical-point issue of Eq. (11)
Thank you for this comment. We would like to clarify that Eq. (11) has a theoretical critical point when the model ability Aˉ\bar{A} equals 0.5. However, this critical point occurs only when Aˉ\bar{A} is exactly 0.5, and any slight deviation (e.g., 0.501) from this value preserves the discriminative ability of instance scores. In our practical implementation, when Aˉ=0.5\bar{A}=0.5, we adopt the original S-based ranking strategy for instance selection. As newly annotated instances are incorporated into training, the teacher's ability evolves and crosses this critical point. As shown in Table 8, the model ability does not exactly equal 0.5 and successfully crosses this point. For mathematical completeness, we added an explicit formulation of Eq. (11) at the critical point.

Furthermore, Eq. (11) provides a simple curriculum-like strategy, favoring easy-to-learn instances when model is weak and challenging ones as it is strong. We further compare it with an alternative strategy, as shown in the table. The close performance suggests that the specific formulation of Eq. (11) has limited impact, as long as it prevents excessive hard instances from being introduced when model is weak. More importantly, the key component of Eq. (11) lies in the ability-aware instance value SS defined in Eq. (10), which adaptively adjusts the weights of different evaluation dimensions according to the model's current ability, enabling the selection of instances that better improve the corresponding abilities. Replacing Eq. (10) with a direct summation of evaluation dimensions (common practice, i.e., ∑Ui\sum U_i) results in a 2.97% mAP drop, showing the importance of our model-state-aware instance value.

StrategySelection FormulamAP (%)
Alternativeexp(−(S−Aˉ)2)exp(-(S-\bar A)^2)35.68
OursAˉS+(1−Aˉ)(1−S)\bar AS+(1-\bar A)(1-S)35.79

Q2: Comparison with ALS Teacher
We have added a comparison with ALS Teacher. As shown in Table, since the original ALS Teacher cannot be directly applied to SAOOD task, we implement its active learning strategy and use the same S2Teacher for pseudo-label mining to ensure a fair comparison. Under 1% annotation ratio, ALS Teacher achieves 30.76% mAP, while Active-SAOOD achieves 35.79% mAP. This performance gap mainly stems from the image-level active learning paradigm adopted by ALS Teacher, which is not well suited to SAOOD. More results will be provided in the final version.

MethodmAP (%)
ALS Teacher30.76
Active-SAOOD35.79

Specifically, ALS Teacher is designed for semi-supervised OOD and selects the highest-scoring images based on image-level sampling metrics. Its sampling strategy therefore follows the similar image-level active learning paradigm (already compared). This paradigm is unsuitable for SAOOD for two reasons: (1) image-level sampling metrics do not match the instance-level selection requirement. (2) directly selecting the most difficult samples without considering the model's current learning state may introduce excessive challenging instances. In contrast, our Active-SAOOD introduces newly instance-level evaluation dimensions tailored to SAOOD and remote sensing, together with a model-state observation strategy that selects suitable instances according to model's current state. This enables instance valuation to consider not only instance information but also the model's abilities and learning needs, enabling the selection of instances that better improve corresponding abilities.

Q3: Whether the gains Δ\Delta in Table 1 come from active selection
For consistent comparison, Δ\Delta in Table 1 represents the performance difference between each method and the baseline Rotated FCOS, rather than the gains solely from active selection. The gains from active selection should be calculated by subtracting the performance of corresponding SAOOD method. Under 1% annotation, Active-SAOOD achieves gains of 5.06% (35.79 − 30.73), 15.35% (38.54 − 23.19), and 10.62% (32.24 − 21.62) mAP over S2Teacher, RSST, and DML, respectively. These positive gains across different SAOOD frameworks demonstrate the effectiveness of active selection. We will revise the relevant descriptions to better distinguish gains from active learning.

Q4: Other minor issues
Thank you for your suggestions. We selected DOTA-v1.0 and DOTA-v1.5 mainly to follow existing SAOOD methods [S2Teacher, RSST] and ensure fair comparisons. We will carefully refine Figs. 1 and 2 and correct the dataset name from DIOR to DIOR-R for greater clarity. Since baseline method without our active learning may suffer from training failures and extremely low performance under 1% annotation ratio, reporting average results will underestimate baseline performance and exaggerate the gains of Active-SAOOD due to training instability. We consistently report max results in Table 1 for fair comparison, while providing more detailed results from repeated experiments in Table 4.

Dear reviewer wa9z, this is a gentle reminder regarding our rebuttal. We would be very grateful if you could take a moment to review our responses when you have time. Thank you again for your time and effort.

Reviewer_wa9zOpenReview ↗

Official comment of wa9z

Thanks to the authors for the thorough response. The authors have addressed some of my concerns with explanation and additional experiments results. I will take these into further consideration. I am inclined to raise my rating to 3.

Thank you very much for your reconsideration and for being inclined to raise your rating to 3. We would appreciate it if you could update your rating in the review system. Thank you again for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.