acceptodds
Under review as a conference paper at ICLR 2027

Semantic Prior-Guided Learning with Adaptive Pseudo-Labeling for Semi-Supervised Oriented Object Detection

Abstract

Semi-supervised oriented object detection provides a promising solution for reducing annotation costs in remote-sensing imagery. However, existing approaches still face two critical challenges: insufficient representation capability caused by limited labeled samples and unreliable pseudo-labels generated from unlabeled data. These challenges are further exacerbated by arbitrary object orientations, dense object distributions, substantial scale variations, and complex inter-class similarities in remote-sensing imagery. To address these issues, we propose SPAL-Net, a semantic prior-guided learning framework with adaptive pseudo-labeling for semi-supervised oriented object detection. SPAL-Net leverages CLIP-derived text embeddings as class-level semantic priors to enhance visual representation learning. Specifically, a Semantic Alignment Module (SAM) is designed to project multi-scale visual features into the text embedding space and enhance visual representations with class-aware semantic information. Furthermore, an Instance-Level Distribution Alignment Loss (IDAL) is proposed to provide semantic supervision by aligning the visual class distribution of each positive instance with the corresponding text-induced semantic distribution. This enables the model to exploit inter-class semantic relationships beyond conventional one-hot supervision, thereby improving the discriminative capability of instance representations. To improve the quality of unlabeled supervision, we develop a Class-Adaptive Pseudo-Label Filtering (CAPF) strategy. By maintaining class-specific confidence thresholds using exponential moving averages, CAPF dynamically adjusts pseudo-label selection based on evolving class-wise confidence statistics, effectively filtering out noisy pseudo-labels and improving the reliability of unlabeled supervision. Extensive experiments on benchmark remote-sensing datasets demonstrate that SPAL-Net achieves state-of-the-art performance under various partial-label settings, validating its effectiveness.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Semi-supervised learning, oriented object detection, adaptive pseudo-labeling, Semantic Alignment

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2iyHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SPAL-Net, a semantic prior-guided framework for semi-supervised oriented object detection. It introduces a Semantic Alignment Module to enhance multi-scale visual features using CLIP text embeddings, an Instance-Level Distribution Alignment Loss to incorporate inter-class semantic relationships, and Class-Adaptive Pseudo-Label Filtering to improve unlabeled supervision. Experiments on DOTA-v1.0 and DOTA-v1.5 demonstrate competitive performance across different annotation budgets.

Critical strengths
  1. The paper explores the integration of vision-language semantic priors into semi-supervised oriented object detection, providing a potentially useful direction for reducing annotation dependence.

  2. The proposed framework incorporates textual priors into both multi-scale feature learning and instance-level supervision, enabling semantic information to influence different stages of detection.

  3. The method achieves strong performance on DOTA-v1.0 and DOTA-v1.5 across multiple annotation budgets, with consistent improvements over the underlying teacher-student baseline.

Critical weaknesses
  1. The paper explores the integration of vision-language semantic priors into semi-supervised oriented object detection, providing a potentially useful direction for reducing annotation dependence.

  2. The proposed framework incorporates textual priors into both multi-scale feature learning and instance-level supervision, enabling semantic information to influence different stages of detection.

Questions for authors
  1. The methodological novelty is relatively limited, as the framework mainly combines CLIP-based feature enhancement, semantic distribution alignment, and class-adaptive pseudo-label filtering within an established Mean Teacher architecture. The paper does not sufficiently distinguish its contributions from existing vision-language distillation and adaptive pseudo-labeling methods.

  2. IDAL assumes that semantic similarities between CLIP class-name embeddings provide appropriate supervision for visual instance discrimination. However, linguistic similarity does not necessarily correspond to visual similarity in aerial imagery, and the paper provides no evidence that the resulting class relationships reflect actual visual confusion patterns.

  3. The effectiveness of CAPF in improving pseudo-label reliability is insufficiently demonstrated. Its thresholds are derived from the mean confidence of predictions that already pass a global threshold, but higher confidence does not necessarily indicate higher pseudo-label precision. The experiments do not directly evaluate pseudo-label precision, recall, or class-wise selection quality against fixed-threshold alternatives.

  4. The ablation studies do not adequately establish whether the improvements originate from meaningful textual semantics rather than additional model capacity or generic regularization. Comparisons using randomized class embeddings, shuffled semantic relationships, or conventional label smoothing are necessary to validate the claimed benefits of semantic prior-guided learning.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2iyH [Part 1/3]

We thank you for your insightful and constructive feedback. Our responses follow.


Question 1:

Methodological novelty of SPAL-Net and its differences from vision–language distillation and adaptive pseudo-labeling methods

Response:

SPAL-Net's central innovation is the cross-stage use of class-level semantic priors: it enhances multi-scale representations before region proposal generation, models class relationships at the RoI level, and links class-adaptive pseudo-label selection to semantic supervision on unlabeled data.

(1) SAM. Unlike vision–language distillation constraints applied to region-level representations, SAM introduces class-level semantic priors into feature learning before region proposal generation, allowing them to inform both region proposals and subsequent RoI representations. Specifically, SAM constructs spatially structured semantic context from dense class responses and modulates multi-scale features before proposal generation, enabling class priors to contribute to proposal generation and instance-level learning.

(2) IDAL. The key difference between IDAL and uniform label smoothing lies in the class relationships encoded by their supervision targets. IDAL uses textual semantic relationships to construct class-dependent auxiliary targets, retaining standard classification supervision while adding class relationships not explicitly represented by one-hot labels. This complements SAM’s feature-level semantic priors with instance-level relational supervision.

(3) CAPF. CAPF shares the motivation of adaptive-threshold methods but differs in the source and update of its threshold statistics. FlexMatch [1] adjusts thresholds according to class learning progress, FreeMatch [2] adaptively adjusts global and class-specific thresholds, and Li et al. [3] jointly consider classification and localization quality. Specifically, CAPF uses detection confidences that pass a fixed global threshold to dynamically update class-specific filtering criteria, balancing a global confidence floor with differences in confidence across classes. In SPAL-Net, the retained pseudo-labels also supervise unlabeled detection and assign class-relationship targets for IDAL.

We have added “Adaptive Pseudo-Label Selection in Semi-Supervised Learning” to Related Work to compare CAPF with prior thresholding methods. Tables 3–5 in Section 4.2 evaluate module combinations, class embeddings, and supervision targets; the latter two comparisons are detailed in Question 4 below.


Question 2:

Whether CLIP text-semantic similarity reflects visual confusion among remote-sensing objects

Response:

We thank the reviewer for raising this question about the rationale for IDAL’s semantic supervision. We do not assume that semantic similarity between CLIP class-name embeddings fully corresponds to visual similarity or class-confusion patterns in remote-sensing images. IDAL uses textual semantic relationships as an auxiliary class prior and applies a KL-divergence constraint to the visual class distribution of positive RoIs. The standard detection classification loss based on ground-truth or pseudo-labels remains responsible for learning visually discriminative features.

To analyze the correspondence between textual semantic relationships and visual confusion, we have added a quantitative analysis and visualization to Appendix E of the revised manuscript (Figure 4). Specifically, we use a baseline teacher model without SAM or IDAL and construct a class-agnostic confusion matrix on DOTA-v1.5 under the 10% labeled-data setting, using matching at rotated IoU ≥ 0.5. Across 120 class pairs, CLIP text similarity has a moderate positive correlation (Spearman ρ = 0.468, permutation-test p = 0.0001) with the bidirectional average visual confusion rate. At confidence thresholds of 0.3 and 0.6, the corresponding correlation coefficients are 0.440 and 0.443, respectively. These results indicate that textual semantic relationships are associated with visual confusion to some extent, but do not fully characterize class confusion in remote-sensing images, particularly its directionality.

In addition, Table 5 in Section 4.2 of the revised manuscript compares the uniform IDAL target with the CLIP-semantic target. With SAM and CAPF disabled under the same setting, the CLIP-semantic target achieves 51.24% mAP, compared with 48.67% for the uniform target, an improvement of 2.57 percentage points. This indicates that non-uniform class relationships provide additional benefit to IDAL.


[1] Bowen Zhang et al. FlexMatch: Boosting Semi-Supervised Learning with Curriculum Pseudo Labeling. NeurIPS, 2021.

[2] Yidong Wang et al. FreeMatch: Self-Adaptive Thresholding for Semi-Supervised Learning. ICLR, 2023.

[3] Hengduo Li, Zuxuan Wu, Abhinav Shrivastava, and Larry S. Davis. Rethinking Pseudo Labels for Semi-Supervised Object Detection. AAAI, 2022.

AuthorsOpenReview ↗

Response to Reviewer 2iyH [Part 2/3]

Question 3:

Whether CAPF improves pseudo-label precision, recall, and class-wise filtering quality

Response:

We thank the reviewer for the suggestions on pseudo-label reliability and its evaluation. Higher confidence does not necessarily imply higher pseudo-label precision. We have therefore added comparisons of pseudo-label quality across filtering strategies and a class-wise analysis to Appendix B of the revised manuscript (Tables 8 and 9). Under the 10% labeled-data setting on DOTA-v1.5, SAM and IDAL are disabled, and teacher models trained with different filtering strategies are evaluated under a common protocol. Table 8 shows that CAPF achieves higher precision and mAP than fixed global and fixed class-specific thresholding, but has lower recall, illustrating the precision–recall trade-off.

Table 8. Pseudo-label quality and detection performance of models trained with different filtering strategies.

Filtering StrategyPrecision (%)Recall (%)mAP (%)
Fixed Global Threshold69.3651.1645.09
Fixed Class-Specific Thresholds74.3350.2647.23
CAPF (Ours)80.5248.6948.78

Table 9 reports selected class-wise results for precision, recall, and the number of retained predictions. CAPF improves precision in 13 of the 16 classes and improves both coverage and precision for Helicopter. However, precision decreases for Soccer ball field and Roundabout, and Container crane still receives no effective coverage, showing that the gains are not uniform across classes.

Table 9. Class-wise pseudo-label quality and retained counts (selected classes).

ClassRecall (%) Global / CAPFPrecision (%) Global / CAPFRetained Global / CAPF
Plane84.41 / 74.1877.51 / 91.114,877 / 3,646
Bridge42.77 / 40.8617.34 / 38.421,944 / 838
Small vehicle41.55 / 40.1271.42 / 80.7742,750 / 36,498
Helicopter0.00 / 35.940.00 / 82.141 / 56
Soccer ball field29.63 / 47.3346.45 / 26.50155 / 434
Roundabout66.20 / 69.3442.89 / 29.31443 / 679
Container crane0.00 / 0.000.00 / 0.006 / 42

AuthorsOpenReview ↗

Response to Reviewer 2iyH [Part 3/3]

Question 4:

Whether the gains from textual semantic priors arise from meaningful semantics rather than additional model capacity or generic regularization

Response:

We thank the reviewer for the valuable suggestions on validating the effectiveness of semantic priors and strengthening the ablation design. Experiments that replace class embeddings and supervision targets support additional benefits from CLIP class representations and their induced relationships. These experiments separately evaluate the contribution of class representations to the overall framework and the contribution of IDAL’s auxiliary supervision targets.

(1) Comparison of different class embeddings.

To determine whether the gains arise from additional model capacity, we compare fixed random, learnable random, and CLIP text embeddings using the same network architecture and training setup under the 10% labeled-data setting on DOTA-v1.5. SAM and IDAL share the corresponding embeddings, while CAPF remains enabled. Table 4 shows that CLIP text embeddings outperform both random alternatives, supporting the contribution of pretrained class representations to the overall framework. The experiment has been added to Section 4.2 of the revised manuscript.

Table 4. Ablation of Class Embeddings.

Embedding TypemAP (%)
Fixed Random Embeddings49.61
Learnable Random Embeddings52.26
CLIP Text Embeddings (Ours)54.77

(2) Comparison of IDAL with conventional label smoothing.

To distinguish IDAL’s gains from those of generic regularization, we compare standard label smoothing, an IDAL variant that retains the auxiliary alignment structure but uses uniform targets, and full IDAL (Table 5) under the 10% labeled-data setting on DOTA-v1.5, with SAM and CAPF disabled. The first method is applied to the RoI classifier, including background samples; the latter two use the same auxiliary structure and differ only in their supervision targets. Full IDAL outperforms both controls, supporting the contributions of the auxiliary alignment structure and the non-uniform semantic targets. This experiment has been added to Section 4.2 of the revised manuscript; module complementarity is discussed in Question 1 above (Table 3).

Table 5. IDAL vs. Label Smoothing.

MethodmAP (%)
Baseline45.09
Standard Label Smoothing45.64
IDAL (Uniform Label Smoothing)48.67
IDAL (Ours)51.24

Thank you again for your review. We hope our response could address your concerns. If you have any further questions, we would be happy to discuss them further.

AuthorsOpenReview ↗

Response to Reviewer 2iyH: Clarification on Ethics Concerns

Response:

We noticed that "Flag For Ethics Review" was marked as "Yes". However, we could not identify any specific ethical concerns in the review comments. The concerns raised primarily relate to methodological novelty, semantic supervision, pseudo-label reliability, and experimental validation, which we have carefully addressed in our rebuttal.

We take research ethics seriously and are committed to adhering to the ICLR Code of Ethics and responsible research practices. To the best of our knowledge, we are not aware of any ethical issues associated with our submission.

We would sincerely appreciate it if the reviewer could clarify any specific ethical concerns regarding our work, so that we can address them appropriately.

Thank you again for your time and consideration.

Reviewer_17GROpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SPAL-Net, a semantic prior-guided framework for semi-supervised oriented object detection in remote sensing imagery. The method addresses two major challenges: insufficient visual representation learning under limited labeled data and unreliable pseudo-labels from unlabeled images. SPAL-Net introduces three components. First, the Semantic Alignment Module (SAM) incorporates CLIP-derived class-level textual priors into multi-scale visual features through semantic response aggregation and gated modulation. Second, the Instance-Level Distribution Alignment Loss (IDAL) aligns visual class distributions with text-induced distributions to exploit semantic relationships among object categories. Third, Class-Adaptive Pseudo-Label Filtering (CAPF) dynamically adjusts class-specific confidence thresholds using exponential moving averages to improve pseudo-label selection. The framework is built upon a teacher-student architecture with Rotated Faster R-CNN. Experiments on DOTA-v1.0 and DOTA-v1.5 demonstrate competitive performance across different annotation budgets. Under the 10% labeled-data setting, SPAL-Net achieves 58.97% and 54.77% mAP on DOTA-v1.0 and DOTA-v1.5, respectively. Component ablation studies further demonstrate the effectiveness of the proposed modules.

Critical strengths

1.The paper addresses two important challenges in semi-supervised oriented object detection by combining semantic representation enhancement with adaptive pseudo-label selection. SAM and IDAL introduce CLIP-derived semantic information at feature and instance levels, while CAPF improves the reliability of unlabeled supervision. The integration is technically reasonable and provides a practical approach to improving detection performance under limited annotation budgets. 2.SPAL-Net achieves competitive results on both DOTA-v1.0 and DOTA-v1.5 across multiple labeling ratios. In particular, the improvements under the 10% annotation setting demonstrate its potential for annotation-efficient remote sensing applications. The ablation results show that CAPF, SAM, and IDAL each contribute to performance improvements, providing useful evidence for the effectiveness of the proposed framework.

Critical weaknesses

1.The main components build on established techniques, including CLIP-based semantic representations, distribution alignment, and EMA-based confidence thresholding. Although their integration is effective, the paper does not sufficiently demonstrate which aspects of the semantic prior are essential for the observed improvements. In particular, comparisons with alternative semantic representations, such as random embeddings, learned class embeddings, or remote-sensing-specific vision-language embeddings, are missing. These comparisons would help establish whether the improvements arise from meaningful transferred semantic knowledge rather than additional model capacity or regularization. 2.Although the reported improvements are promising, the experimental validation has several limitations. The paper does not report results across multiple random seeds or provide measures of statistical variability. Furthermore, the computational overhead introduced by the semantic alignment and distribution alignment components is not quantified. The ablation study also lacks pairwise module combinations, making it difficult to fully assess potential interactions among the three components. These omissions weaken the evidence for the robustness and practical efficiency of the proposed framework.

Questions for authors
  1. How much of the performance improvement can be attributed specifically to the semantic knowledge encoded in CLIP? Could the authors compare the proposed textual priors with random class embeddings, learnable class embeddings, or remote-sensing-specific vision-language representations under the same experimental settings?
  2. Could the authors provide additional ablations for pairwise combinations of the proposed components, particularly SAM + IDAL, SAM + CAPF, and IDAL + CAPF? Such results would help clarify whether the components provide complementary benefits or exhibit substantial interactions.
  3. How sensitive is CAPF to its global confidence threshold and EMA momentum? Could the authors provide a comparison with fixed class-specific thresholds or other adaptive thresholding strategies to better isolate the benefits of the proposed filtering mechanism?
  4. What are the additional training and inference costs introduced by SPAL-Net compared with the teacher-student baseline? Reporting parameter counts, computational complexity, and runtime would help assess its practical applicability.
  5. Are the reported improvements consistent across different random seeds and labeled-data splits? Providing the mean and standard deviation for representative settings would help establish the robustness of the experimental findings.
2 comments
AuthorsOpenReview ↗

Response to Reviewer 17GR [Part 1/2]

We thank you for your insightful and constructive feedback. Our responses follow.


Question 1:

The practical contribution of CLIP semantics and comparison with random class embeddings

Response:

We thank the reviewer for the suggestion to validate the effectiveness of the CLIP semantic prior. Under the 10% labeled-data setting on DOTA-v1.5, we compare fixed random, learnable random, and CLIP text embeddings. SAM and IDAL share the corresponding embeddings, and CAPF remains enabled. CLIP embeddings outperform both random alternatives. Because the embeddings used by SAM and IDAL are replaced together, this experiment primarily evaluates the contribution of semantic priors to the overall framework. The results have been added as Table 4 in Section 4.2 of the revised manuscript.

Table 4. Ablation of Class Embeddings.

Embedding TypemAP (%)
Fixed Random Embeddings49.61
Learnable Random Embeddings52.26
CLIP Text Embeddings (Ours)54.77

Question 2:

Effects and complementarity of pairwise combinations of SAM, IDAL, and CAPF

Response:

We thank the reviewer for the suggestion to make the ablation study more complete. We have added experiments on all pairwise combinations of the three modules under the 10% labeled-data setting on DOTA-v1.0 and DOTA-v1.5, using the same teacher–student framework. Each pairwise combination outperforms the corresponding single-module variants, and the full SPAL-Net further outperforms the best two-module combination, supporting the complementarity of the modules. The complete results have been added as Table 3 in Section 4.2 of the revised manuscript.

Table 3. Component ablation under the 10% annotation protocol.

CAPFSAMIDALDOTA-v1.0DOTA-v1.5
–––49.1645.09
✓––52.1448.78
–✓–55.4051.83
––✓55.4151.24
✓✓–57.4353.61
✓–✓57.9353.46
–✓✓58.0453.87
✓✓✓58.9754.77

Question 3.1:

Sensitivity of CAPF to the global confidence threshold and EMA momentum

Response:

We thank the reviewer for the suggestion to evaluate CAPF’s parameter sensitivity. We have added offline sensitivity analyses to Appendix A of the revised manuscript to assess how the global confidence threshold and EMA momentum affect pseudo-label quality.

Using the same trained CAPF teacher, we evaluate offline how the global threshold δ and EMA momentum ρ affect pseudo-label precision and recall (Tables 6 and 7); no retraining is performed for these parameter settings. Table 6 shows that increasing the global threshold improves precision but reduces recall, illustrating the trade-off between precision and coverage.

Table 6. Sensitivity to δ (ρ = 0.9996).

δPrecision (%)Recall (%)
0.578.0149.43
0.680.5248.69
0.782.1244.09
0.885.2639.85
0.989.1529.87

Table 7 shows that pseudo-label quality is relatively stable across the EMA momentum values considered. We use δ = 0.6 and ρ = 0.9996 as the default configuration to balance pseudo-label reliability and coverage.

Table 7. Sensitivity to ρ (δ = 0.6).

ρPrecision (%)Recall (%)
0.990079.5646.42
0.999080.2548.32
0.999680.5248.69
0.999980.5948.57

AuthorsOpenReview ↗

Response to Reviewer 17GR [Part 2/2]

Question 3.2:

Comparison of CAPF with fixed class-specific thresholds

Response:

We thank the reviewer for the suggestion to compare CAPF with fixed-threshold strategies. To evaluate dynamic class-specific threshold updates, we compared fixed global thresholding, fixed class-specific thresholding, and CAPF. The results of these experiments are reported in Table 8 of Appendix B of the revised manuscript. The three strategies use the same teacher–student framework and are trained separately; their pseudo-label quality and detection performance are then evaluated under the same protocol. Fixed class-specific thresholds outperform the fixed global threshold. Compared with fixed class-specific thresholds, CAPF further improves pseudo-label precision and mAP, with a slight decrease in recall.

Table 8. Pseudo-label quality and detection performance of models trained with different filtering strategies.

Filtering StrategyPrecision (%)Recall (%)mAP (%)
Fixed Global Threshold69.3651.1645.09
Fixed Class-Specific Thresholds74.3350.2647.23
CAPF (Ours)80.5248.6948.78

Question 4:

Parameter count, computational cost, and inference-speed overhead of SPAL-Net relative to the teacher–student baseline

Response:

We thank the reviewer for asking about SPAL-Net’s computational cost and inference efficiency. We have added comparisons of model parameters, FLOPs, and FPS to Appendix C of the revised manuscript (Table 10), with measurements taken on an NVIDIA RTX 3090 GPU. Compared with the baseline, SPAL-Net has more parameters and FLOPs and lower inference speed. The additional inference computation comes primarily from SAM; IDAL and CAPF are used only during training. The parameter count includes the IDAL projection layers but excludes the frozen CLIP text encoder. This comparison quantifies the trade-off between computational overhead and improved detection performance.

Table 10. Model complexity and inference speed.

MethodParameters (M)FLOPs (G)FPS
Baseline41.139211.3024.16 ± 0.17
SPAL-Net42.586235.8120.89 ± 0.24

Question 5:

Performance stability across training random seeds and labeled-data splits

Response:

We thank the reviewer for the suggestion to evaluate experimental robustness. We have added multi-seed experiments to Appendix D of the revised manuscript (Table 11). Under the 10% labeled-data setting on DOTA-v1.5 and using the fixed SOOD split, we train the baseline and SPAL-Net with three random seeds (42, 128, and 110280) and report the mean and standard deviation of mAP. SPAL-Net outperforms the baseline with each seed and exhibits low variation, supporting training stability on the fixed split. Results across different annotation budgets also show consistent gains.

Table 11. Detection performance across training random seeds on DOTA-v1.5 with 10% labeled data using the fixed SOOD split.

MethodSeed 1Seed 2Seed 3Mean ± Std (%)
Baseline45.0945.2145.1245.14 ± 0.06
SPAL-Net54.7754.8254.6454.74 ± 0.09

Thank you again for your review. We hope our response could address your concerns. If you have any further questions, we would be happy to discuss them further.

Reviewer_ZTA3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPAL-Net, a semi-supervised oriented object detection framework that addresses limited visual representations and unreliable pseudo-labels in remote-sensing imagery. The method introduces three components: a Semantic Alignment Module (SAM) to incorporate CLIP-derived semantic priors into multi-scale features, an Instance-Level Distribution Alignment Loss (IDAL) to exploit inter-class semantic relationships, and a Class-Adaptive Pseudo-Label Filtering (CAPF) strategy to improve pseudo-label selection. Experiments on DOTA-v1.0 and DOTA-v1.5 demonstrate competitive performance across different annotation settings, supported by component ablations and temperature sensitivity analyses.

Critical strengths

The paper addresses a practically important problem in semi-supervised oriented object detection, with a clear motivation and a well-structured framework. Its main strength lies in the strong empirical performance across different annotation ratios on DOTA-v1.0 and DOTA-v1.5, demonstrating the practical effectiveness of the proposed approach. Additionally, the integration of semantic priors and adaptive pseudo-label filtering provides a reasonable strategy for improving detection under limited supervision.

Critical weaknesses
  1. The framework largely combines established techniques, including CLIP-based feature enhancement (SAM), KL-based semantic alignment (IDAL), and EMA-based adaptive pseudo-label filtering (CAPF). IDAL resembles semantic label smoothing and knowledge distillation, while CAPF is closely related to existing adaptive thresholding methods [1–3]. The authors should clarify the technical novelty of these components, particularly CAPF, relative to prior work.

  2. CAPF estimates class-specific thresholds only from predictions passing a global threshold, potentially excluding difficult or underrepresented instances. Moreover, when a class first appears with a single prediction, its threshold equals that prediction's confidence, causing it to be discarded under Eq. (12). The authors should clarify these limitations and provide class-wise pseudo-label quality analyses.

  3. IDAL assigns identical semantic targets to all instances of the same class, regardless of their visual characteristics, making it similar to semantic label smoothing. The authors should clarify its instance-level benefits and provide comparisons with conventional soft-label supervision.

  4. Table 3 only evaluates individual components and the full model, without examining pairwise combinations such as SAM+IDAL and SAM+CAPF. More comprehensive ablations are needed to assess the modules' individual contributions and complementarity.

  5. The effectiveness of key components lacks direct validation. CAPF lacks pseudo-label precision, recall, and class-wise coverage analyses, while IDAL lacks comparisons with standard label smoothing and random semantic prototypes. These evaluations are needed to verify whether the gains stem from improved pseudo-label quality and semantic priors.

  6. Figure 3 reports different mAP values for the selected temperature configuration across the two parameter sweeps. The authors should clarify the fixed parameter settings in each experiment and explain this apparent inconsistency to ensure reproducibility.

[1] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. 2021. FlexMatch: Boosting Semi-Supervised Learning with Curriculum Pseudo Labeling. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 34, 18408–18419. Paper. [2] Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, Bernt Schiele, and Xing Xie. 2023. FreeMatch: Self-Adaptive Thresholding for Semi-Supervised Learning. In International Conference on Learning Representations (ICLR). Paper. [3] Hengduo Li, Zuxuan Wu, Abhinav Shrivastava, and Larry S. Davis. 2022. Rethinking Pseudo Labels for Semi-supervised Object Detection. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), Vol. 36, No. 2, 1314–1322. https://doi.org/10.1609/aaai.v36i2.20019.

Questions for authors

Please see the Weaknesses section.

4 comments
AuthorsOpenReview ↗

Response to Reviewer ZTA3 [Part 1/3]

We thank you for your insightful and constructive feedback. Our responses follow.


Question 1:

Technical novelty of SAM, IDAL, and CAPF and their differences from related methods

Response:

The central innovation of SPAL-Net lies in using class-level semantic priors across multiple stages: it enhances multi-scale representations before region proposal generation and models class relationships at the RoI level, while linking class-adaptive pseudo-label selection with semantic supervision on unlabeled data.

(1) SAM. Unlike vision–language distillation constraints applied to region-level representations, SAM introduces class-level semantic priors into feature learning before region proposal generation, allowing them to inform both region proposals and subsequent RoI representations. Specifically, SAM constructs spatially structured semantic context from dense class responses and modulates multi-scale features before proposal generation, enabling class priors to contribute to proposal generation and instance-level learning.

(2) IDAL. The key difference between IDAL and uniform label smoothing lies in the class relationships encoded by their supervision targets. IDAL uses textual semantic relationships to construct class-dependent auxiliary targets, retaining standard classification supervision while adding class relationships not explicitly represented by one-hot labels.

(3) CAPF. CAPF shares the general motivation of existing adaptive-threshold methods, but differs in the source of its threshold statistics and how they are updated. FlexMatch [1] adjusts thresholds according to class learning progress, FreeMatch [2] adaptively adjusts global and class-specific thresholds, and Li et al. [3] jointly consider classification and localization quality. Specifically, CAPF uses detection confidences that pass a fixed global threshold to dynamically update class-specific filtering criteria, balancing a global confidence floor with differences in confidence across classes.

We have added “Adaptive Pseudo-Label Selection in Semi-Supervised Learning” to Related Work to discuss these distinctions.


Question 2.1:

Effects of CAPF’s global pre-filter on difficult or rare classes and class-wise pseudo-label quality

Response:

CAPF adapts to confidence differences across classes and training stages to improve pseudo-label reliability. However, its global pre-filter may discard valid low-confidence predictions, reducing recall for difficult or rare classes.

Under the 10% labeled-data setting on DOTA-v1.5, we evaluated teacher models trained with different filtering strategies under a common protocol. The detailed results are reported in Tables 8 and 9 in Appendix B of the revised manuscript and are reproduced below for convenience. CAPF achieves higher pseudo-label precision and mAP than the two fixed-threshold strategies, but has lower recall, reflecting a trade-off between pseudo-label reliability and coverage.

Table 8. Pseudo-label quality and detection performance of models trained with different filtering strategies.

Filtering StrategyPrecision (%)Recall (%)mAP (%)
Fixed Global Threshold69.3651.1645.09
Fixed Class-Specific Thresholds74.3350.2647.23
CAPF (Ours)80.5248.6948.78

Table 9 gives class-wise results. CAPF improves precision in 13 of 16 classes and Helicopter recall, but reduces recall for some classes and provides no effective coverage for Container crane. Its gains vary by class and do not resolve limited coverage for difficult and rare classes.

Table 9. Class-wise pseudo-label quality and retained counts (selected classes).

ClassRecall (%) Global / CAPFPrecision (%) Global / CAPFRetained Global / CAPF
Plane84.41 / 74.1877.51 / 91.114,877 / 3,646
Bridge42.77 / 40.8617.34 / 38.421,944 / 838
Small vehicle41.55 / 40.1271.42 / 80.7742,750 / 36,498
Helicopter0.00 / 35.940.00 / 82.141 / 56
Container crane0.00 / 0.000.00 / 0.006 / 42

[1] Bowen Zhang et al. FlexMatch: Boosting Semi-Supervised Learning with Curriculum Pseudo Labeling. NeurIPS, 2021.

[2] Yidong Wang et al. FreeMatch: Self-Adaptive Thresholding for Semi-Supervised Learning. ICLR, 2023.

[3] Hengduo Li, Zuxuan Wu, Abhinav Shrivastava, and Larry S. Davis. Rethinking Pseudo Labels for Semi-Supervised Object Detection. AAAI, 2022.

AuthorsOpenReview ↗

Response to Reviewer ZTA3 [Part 2/3]

Question 2.2:

Threshold initialization when a class first appears

Response:

We thank the reviewer for pointing out this edge case.

When a class first appears with only one prediction passing the global pre-filter, its initial threshold equals that prediction’s confidence. The strict greater-than condition therefore discards it, potentially limiting early class coverage. We clarify this edge case after Equation (12) in Section 3.4.


Question 3:

IDAL’s instance-level supervision mechanism and its advantages over conventional soft-label supervision

Response:

We thank the reviewer for the helpful comments on IDAL’s instance-level semantic supervision. We further clarify what makes IDAL instance-level and how it differs from conventional label smoothing.

(1) Instance-level semantic supervision in IDAL.

IDAL is instance-level in that it constructs and aligns visual distributions independently for each instance; it does not assign different semantic targets to instances of the same class. Each positive RoI produces its own visual distribution, piv=Softmax⁡(viT⊤/τv)p_i^v=\operatorname{Softmax}(v_iT^\top/\tau_v), while its semantic target is selected by class from Qt=Softmax⁡(TT⊤/τt)Q^t=\operatorname{Softmax}(TT^\top/\tau_t), namely qit=Qyi,:tq_i^t=Q^t_{y_i,:}. Thus, instances of the same class share a target, but the visual distributions and their KL-divergence constraints are computed separately for each RoI (Equations (6)–(8)).

(2) Comparison with conventional soft-label supervision.

Unlike conventional label smoothing, IDAL uses CLIP text semantics to construct non-uniform class targets. To evaluate its effect, we have added a controlled comparison to Table 5 in Section 4.2 of the revised manuscript. Under the 10% labeled-data setting on DOTA-v1.5, with SAM and CAPF disabled, we compare standard label smoothing applied directly to the RoI classifier, an IDAL variant that retains the auxiliary alignment structure but uses uniform smoothing targets, and the full IDAL. Full IDAL outperforms both controls, improving mAP by 2.57 percentage points over the uniform-target variant. This supports the additional value of non-uniform semantic relationships.

Table 5. IDAL vs. Label Smoothing.

MethodmAP (%)
Baseline45.09
Standard Label Smoothing45.64
IDAL (Uniform Label Smoothing)48.67
IDAL (Ours)51.24

Question 4:

Pairwise ablations of SAM, IDAL, and CAPF and the complementarity of the modules

Response:

We thank the reviewer for the suggestion to make the ablation study more complete. We have added experiments on all pairwise combinations of the three modules under the 10% labeled-data setting on DOTA-v1.0 and DOTA-v1.5, using the same teacher–student framework. Each pairwise combination outperforms the corresponding single-module variants, and the full SPAL-Net further outperforms the best two-module combination, supporting the complementarity of the modules. The complete results have been added as Table 3 in Section 4.2 of the revised manuscript.

Table 3. Component ablation under the 10% annotation protocol.

CAPFSAMIDALDOTA-v1.0DOTA-v1.5
–––49.1645.09
✓––52.1448.78
–✓–55.4051.83
––✓55.4151.24
✓✓–57.4353.61
✓–✓57.9353.46
–✓✓58.0453.87
✓✓✓58.9754.77

AuthorsOpenReview ↗

Response to Reviewer ZTA3 [Part 3/3]

Question 5:

CAPF’s precision, recall, and class-wise coverage, and comparisons of IDAL with label smoothing and random semantic prototypes

Response:

We thank the reviewer for noting that the effectiveness of the key components requires more direct experimental validation. In response, we have added an analysis of teacher prediction quality, class-embedding replacement experiments, and comparisons of IDAL supervision targets to evaluate CAPF and the contribution of textual semantic priors.

(1) Pseudo-label quality and class-wise coverage of CAPF.

We conducted supplementary experiments under the 10% labeled-data setting on DOTA-v1.5. The detailed results are reported in Tables 8 and 9 in Appendix B of the revised manuscript. Teacher models trained with different filtering strategies were evaluated under a common protocol. Table 8 shows that CAPF achieves higher pseudo-label precision and mAP than the fixed-threshold alternatives, but has lower recall. The class-wise results in Table 9 show that CAPF improves precision in most classes, although the gains vary across classes.

(2) Comparison of IDAL with conventional label smoothing and random semantic embeddings.

We have added two supplementary ablation studies to Section 4.2 of the revised manuscript, with the results reported in Tables 4 and 5. Table 4 compares fixed random, learnable random, and CLIP text embeddings, showing that CLIP embeddings perform best. Table 5 compares supervision targets with SAM and CAPF disabled and the IDAL auxiliary structure retained. The results show that CLIP-induced non-uniform class relationships provide additional benefits.

Table 4. Ablation of Class Embeddings.

Embedding TypemAP (%)
Fixed Random Embeddings49.61
Learnable Random Embeddings52.26
CLIP Text Embeddings (Ours)54.77

Question 6:

Fixed settings and differences in results between the two temperature sweeps in Figure 3

Response:

We thank the reviewer for noting that the parameter settings in the temperature sensitivity experiments in Figure 3 were not sufficiently clear. In the visual-temperature sweep, the text temperature is fixed at τₜ = 0.01; in the text-temperature sweep, the visual temperature is fixed at τᵥ = 0.3. Thus, the setting (τᵥ, τₜ) = (0.3, 0.01) in the left plot is the same as τₜ = 0.01 in the right plot, and both yield 51.08% mAP. The default setting in the right plot, (0.3, 0.1), yields 51.24% mAP. We use τᵥ = 0.3 and τₜ = 0.1 as the default values. We have specified the fixed parameter in each sweep in the caption of Figure 3 and in Section 4.2 of the revised manuscript to remove ambiguity and facilitate reproduction.


Thank you again for your review. We hope our response could address your concerns. If you have any further questions, we would be happy to discuss them further.

Thank you for your detailed responses and additional experiments. Most of my concerns have been addressed, and I have therefore increased my score. I appreciate the authors' efforts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.