PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection

Q5. The hyperparameters (radius and $\alpha$).

A5. We appreciate the concern about the sensitivity of thresholds. To address this, we conducted additional experiments:

1. Using the same thresholds as those applied to the DOTA dataset, we verified our method on the entirely different SKU110K-R retail dataset. Our method achieved a mAP$_{50}$ of 62.76, significantly outperforming Point2RBox's mAP$_{50}$ of 3.41. The poor performance of Point2RBox can be attributed to the large inter-class variability in SKU110K-R, which makes it challenging to generate suitable patterns for class representation. Additionally, PointOBB's SSC loss enforces inter-class consistency constraints, which inherently limits its applicability to single-class datasets like SKU110K-R. In contrast, our method remains effective in such scenarios, demonstrating its robustness and lower sensitivity to threshold parameters.

2. We conducted experiments on the DOTA dataset with varying parameters and observed minimal fluctuations in the final results. Additionally, the performance consistently exceeded that of PointOBB by a significant margin (about 8% gain on Oriented RCNN). Therefore, we conclude that the introduced thresholds do not require precise tuning to achieve strong performance, indicating that the model is not highly sensitive to threshold settings.

Q1. The intuition behind the label assignment strategy.

A1. Our label assignment strategy is designed to ensure high certainty for both positive and negative samples. This means that samples classified as positive or negative have a high probability of being correctly labeled, while the "ignored" regions represent areas where positive and negative samples overlap. During training, the model tends to predict the entire object region up to its boundaries as the corresponding class in the CPM to minimize the loss. This allows the model to learn object contours instead of circular CPM patterns, despite differences between the training targets and predictions. The "ignored" label plays a crucial role in this, preventing the model from learning circular patterns. Regarding the hyperparameters in label assignment, we set the radius for positive samples as small as possible. For α\alphaα, we set it to 1, directly using the smallest radius for the circular assignment.

Q2. The advantages of modular approaches' flexibility.

A2. In the task of point-supervised oriented object detection, end-to-end methods like Point2RBox do not generate pseudo-labels directly; instead, they train the model to produce final outputs directly. In contrast, our method adopts a single-branch design that separates pseudo-label generation from detector training. While this approach does not support dynamic updating of pseudo-labels, it offers significant advantages in terms of flexibility. By decoupling pseudo-label generation, our method allows these labels to be reused across different detectors. This modularity enables seamless integration of newer or improved detectors without modifying the overall framework. On the other hand, end-to-end methods embed the detector within the model, making it difficult to replace or adapt, which greatly reduces their flexibility and limits experimental exploration.

Additionally, we would like to clarify a potential misunderstanding regarding dynamic pseudo-label updates. Unlike semi-supervised object detection, weakly-supervised oriented object detection paradigms—such as ours—do not involve dynamic pseudo-label updates during training. In semi-supervised frameworks, dynamic pseudo-label updates are feasible because a small portion of the data includes accurate RBox annotations, allowing the teacher model to generate pseudo-labels with high accuracy from the outset. However, in point-supervised tasks, such as ours, only point annotations are available, and the lack of initial RBox annotations makes dynamic pseudo-label updates inapplicable. This distinction highlights the fundamental differences between semi-supervised and weakly-supervised frameworks.

Q3. Why higher probabilities surrounding object centers and along their axes.

A3. Thank you for pointing out this question. We explain this phenomenon in detail as follows:

• Surrounding object centers: Positive samples are assigned within a small circle around the ground truth center, enabling the model to learn higher probabilities in these regions.
• Along their axes: This characteristic often occurs with symmetric objects, such as sports courts, airplanes, and vehicles. In these cases, the axes exhibit less uncertainty when the model outputs high probabilities, which may be akin to symmetric learning. This suggests that the model is effectively capturing the inherent symmetry in object orientations.

Q4. What is the term 'Ignored' in Figure 2.

A4. "Ignored" indicates that these labels are excluded from training. This means that no matter what the CPM outputs at these points, they do not contribute to the loss calculation. This approach helps the model focus on regions with clear label assignments while avoiding ambiguous areas.

Dear Reviewer oGEA,

Could you kindly review the rebuttal thoroughly and let us know whether the authors have adequately addressed the issues raised or if you have any further questions.

Best,

AC of Submission1151

Dear Reviewers,

May we kindly ask if our responses have addressed your concerns? We look forward to further discussions and feedback from you!

Sincerely,

Authors

Thanks to the authors for responding to most of my questions and the listed weakness. However, I notice that Q1 remains unanswered, and I've been waiting for a feedback on Q1.

Q1. The proposed method (i) is a simple increment to PointOBB.

A1. We apologize if the paper has caused any misunderstanding. Here, we would like to provide further clarification. PointOBB is based on multiple instance learning (MIL) and symmetric learning, while PointOBB-v2 relies on class probability maps (CPM) and principal component analysis (PCA). Therefore, PointOBB and PointOBB-v2 follow entirely different technical paradigms. Specifically，unlike PointOBB, which uses a computationally expensive dual-branch structure and enforces consistency across various image transformations, our method utilizes a streamlined single-branch design. By leveraging a carefully crafted positive-negative sample assignment strategy, we train the CPM and subsequently use weighted PCA to infer the orientation. This innovative approach is both simple and effective, leading to improved detection accuracy while significantly reducing time and memory consumption.

Moreover, some reviewers have acknowledged our novelty and highlighted that our work introduces an innovative pipeline compared to PointOBB. For example, reviewer ND69 mentions that "PointOBB-v2 introduces a novel pipeline," reviewer R8EQ states that "This paper presents PointOBB-v2, a novel approach for single point supervised oriented object detection, ..., providing a lightweight approach that avoids the need for prior deep network components or cumbersome teacher-student training frameworks," and reviewer oGEA highlights that "the authors introduce PointOBB-v2, which presents a streamlined pipeline that eliminates the traditional teacher-student architecture."

Thus, we believe this work represents a meaningful contribution, rather than a marginal improvement.

Q2. (ii) involves many thresholds, indicating that it might not be an easily generalizable one.

A2. We appreciate the concern about the sensitivity of thresholds. To address this, we conducted additional experiments:

1. Using the same thresholds as those applied to the DOTA dataset, we verified our method on the entirely different SKU110K-R retail dataset. Our method achieved a mAP$_{50}$ of 62.76, significantly outperforming Point2RBox (mAP$_{50}$ of 3.41). The poor performance of Point2RBox can be attributed to the large inter-class variability in SKU110K-R, which makes it challenging to generate suitable patterns for class representation. Additionally, PointOBB's SSC loss enforces inter-class consistency constraints, which inherently limits its applicability to single-class datasets like SKU110K-R. In contrast, our method remains effective in such scenarios, demonstrating its robustness and lower sensitivity to threshold parameters.

2. We conducted experiments on the DOTA dataset, focusing on parameters that could potentially impact the results ($\alpha$, b1, b2), and observed minimal fluctuations in the final results. Additionally, the performance consistently exceeded that of PointOBB (33.31%) by a significant margin (minimum 7% gain on Oriented RCNN). Therefore, we conclude that the introduced thresholds do not require precise tuning to achieve strong performance, indicating that the model is not highly sensitive to threshold settings.

Q3. The paper shows weakly supervised detection but lags in performance (45 vs 70+) without cost-performance discussion.

A3. Thank you for raising this important question. We agree that it is necessary to discuss the trade-offs between full annotation cost and performance degradation, and we provide the following explanations:

1. In the field of object detection, the cost of point annotations is about 36.5% lower than HBB and 104.8% lower than OBB, and the efficiency of labeling point is significantly higher than both HBB and OBB, refer to google annotation platform pricing. This cost analysis has also been adopted and validated in prior works, such as H2RBox，PointOBB and Point2RBox, further demonstrating its reliability and relevance.

2. Although the performance of point supervision still lags behind that of full supervision, our method takes a significant step forward compared to existing approaches such as PointOBB, achieving improvements of 10.90/25.15/21.19 mAP$_{50}$ on the DOTA-v1.0/1.5/2.0 datasets. We believe that future research can further narrow the gap, and this work provides a strong example to build upon.

Q4. Regarding the suitability of the work for ICLR (representation learning).

A4. We believe that our submitted work aligns with the scope of ICLR for the following reasons:

1. ICLR specifies the following primary areas for submission: "applications to computer vision, audio, language, and other modalities" and "unsupervised, self-supervised, semi-supervised, and supervised representation learning." We consider our work to be consistent with both these areas.

2. Weakly supervised oriented object detection algorithm H2RBox [1] and fully supervised oriented object detection algorithm KFIoU [2] were both accepted at ICLR 2023, demonstrating the relevance of oriented object detection to ICLR.

3. This year's ICLR submissions also include a significant number of papers related to oriented and general object detection.

[1] H2RBox: Horizontal Box Annotation is All You Need for Oriented Object Detection, ICLR, 2023

[2] The KFIoU Loss for Rotated Object Detection, ICLR, 2023

Q5. How does PointOBBv2 perform on DIOR-R and other oriented object detection datasets?

A5. To further validate the generalizability of our method, we conducted experiments on DIOR-R and SKU110-R. The results, presented in following table, demonstrate that our method consistently improves performance on datasets with diverse object categories and distributions. This indicates the broader applicability and robustness of our approach beyond the DOTA dataset.

Dear Reviewer TgZt,

Could you kindly review the rebuttal thoroughly and let us know whether the authors have adequately addressed the issues raised or if you have any further questions.

Best,

AC of Submission1151

Dear authors,

Thank you for addressing all my questions and concerns. I will re-assess. I do not have any further questions.

Dear Reviewers,

May we kindly ask if our responses have addressed your concerns? We look forward to further discussions and feedback from you!

Sincerely,

Authors

Q1. The generalization of PointOBB-v2 to other domains.

A1. Thank you for pointing out the need to demonstrate broader applicability. As the following table shows, we conducted experiments in the retail scene using the SKU110-R dataset, where our method achieved an mAP$_{50}$ of 62.76. This result highlights the generalizability of PointOBB-v2 to domains beyond aerial imagery. In future work, we plan to further explore its application in other areas, such as autonomous vehicles and medical imaging.

Q2. Grammatical issues and writing.

A2. We appreciate your constructive feedback on improving the writing quality. We will revise the paper to address the grammatical issues you highlighted. Specifically, we will rephrase "Diverse from those approaches relying on one-shot samples" to "Unlike approaches that rely on one-shot samples" and adjust "achieving significant improvements in the accuracy and generation speed of the pseudo-label, and improving memory efficiency" to "achieving significant improvements in pseudo-label accuracy, generation speed, and memory efficiency." Additionally, we will thoroughly review the manuscript to ensure grammatical consistency.

Q3. How model contribute to high-density scenarios.

A3. Our method excels in high-density object scenarios primarily due to two key components:

• Label assignment strategy (L259–266): This ensures accurate and efficient sampling of positive and negative labels, even in dense object distributions.
• Vector constraint (L309–323): This encourages precise orientation prediction while maintaining robustness in cluttered scenes. These components collectively enhance the model's performance in high-density scenarios by addressing the unique challenges of such environments.

Q4. If any alternative methods for estimating orientation and boundary.

A4. Principal Component Analysis (PCA) was chosen for its simplicity, efficiency, and compatibility with weakly supervised learning. While other methods, such as the symmetric learning introduced in H2RBox-v2 [1], could potentially provide better orientation estimation, they may also introduce additional training complexity and time. Clustering algorithms might help in estimating boundaries, and we are considering exploring this direction in future work to further enhance the accuracy and flexibility of our method.

[1] H2RBox-v2: Incorporating Symmetry for Boosting Horizontal Box Supervised Oriented Object Detection, NeurIPS, 2023

Q1. Whether the resolution while learning the CPM affects the performance.

A1. Thank you for raising this point. We use the highest resolution feature map of the FPN to ensure finer representation of orientation and boundary details. For example, a single pixel in the highest resolution CPM corresponds to $4 \times 4 = 16$ pixels in the original image. This level of detail is crucial because many object annotations in DOTA-v1.5 and DOTA-v2.0 are smaller than 10 pixels. Using the highest resolution feature map allows us to better capture these small objects, ensuring more precise orientation and boundary predictions.

Q2. Some typographical inconsistencies.

A2. We appreciate your attention to detail. We will address the typographical issues in Figure 2, Line 433, and Line 468, as well as ensure consistency in equation punctuation throughout the revised manuscript.

Q3. The "randomness" in Line 288.

A3. The "randomness" refers to the variance introduced during the PCA computation when sampling points based on their probabilities. Directly sampling points according to their pixel-level probabilities can result in variations across different passes, potentially disturbing the orientation estimation. To mitigate this, we compute a "weighted PCA," where probabilities are treated as weights rather than relying on stochastic sampling. This approach reduces variance and provides a more stable estimation of object orientation.

Q4. The more pronounced improvements on DOTA-v1.5/v2.0.

A4. The greater improvement on DOTA-v1.5 and DOTA-v2.0 can be attributed to the higher density of objects per image in these datasets. In such dense scenarios, methods like PointOBB face challenges due to the need to constrain the number of RoI proposals, leading to performance drops. Our approach, on the other hand, is robust to dense object distributions, allowing it to effectively handle the increased complexity of these datasets. This robustness underscores the strengths of our method in challenging scenarios.

Q1. Regarding the unclear explanation of CPM.

A1. We apologize for the lack of clarity in our original submission and provide a detailed explanation as follows:

• Training details, loss function of CPM We appreciate your feedback and would like to clarify the training process. The CPM training details are introduced in L363–365: we train the model for 6 epochs using momentum SGD as the optimizer, with a weight decay of 1e-4. The initial learning rate is set to 0.005 and decays by a factor of 10 after the 4$^{th}$ epoch. The activation function used in the middle layers is LeakyReLU, while the last layer employs a sigmoid function. For each point, we first generate its target using a label assignment strategy. Subsequently, we train the CPM using focal loss.
• Ignored labels processing: Ignored labels are excluded from the training process to ensure focus on well-defined samples.
• Structure order: We apologize for the lack of clarity in the structure of the Method section. We will revise the paper to reorder the subsections and provide a smoother explanation. Specifically, we plan to first describe the model training process and then detail the label assignment strategy to improve the flow and understanding.

Q2. How CPM works.

A2. Thank you for your insightful feedback. We provide the following clarifications:

• Why the CPM learns object contours: Our label assignment strategy does not rely solely on a fixed radius for sample assignment. While the positive samples are assigned using a fixed radius, the negative samples are determined using a variable radius strategy (L251–257). Specifically, for each GT object, we calculate the radius based on the distance to its nearest neighbor GT. This results in N radii, which are used to draw circles for negative sample assignment. The area outside these circles is designated as negative samples. Unlike fixed-length assignment, our method adapts to varying object densities, resulting in dynamic sample definitions. We also provide visualization results comparing our dynamic negative radius label assignment with the fixed negative radius approach. We found that when using a fixed radius, the CPM tends to learn object contours that are more circular, with the CPM primarily concentrated around the object's center. In contrast, our method produces clearer object boundaries. Please see the Appendix in our revised paper for detailed visualizations.

  The design of the positive and negative sample definitions ensures high certainty in their classification. Ignored regions, representing areas where positive and negative samples overlap, allow the model to focus on clear distinctions between object and non-object regions. This inherently guides the model to learn object contours, as expanding predictions to the true boundary minimizes the loss. The ignored labels play a critical role in preventing the model from learning circular CPM patterns, thereby enabling contour learning.

• Training target vs. output CPM: The differences between the training target and the output CPM have been addressed above, highlighting how the model effectively learns object contours despite coarse supervision.

• Clarification of Figures 2 and 3: We apologize for the lack of clarity in the figures. In Figure 2, the red point represents the ground truth point annotation. In Figure 3, blue, pink, and orange regions denote the highest CPM values for the classes harbor, ship, and large-vehicle, respectively.

Q3. Regarding generalization to other scenarios.

A3. Indeed, our method performs exceptionally well in dense object scenarios. To further demonstrate this, we conducted additional experiments on the SKU110K-R dataset, achieving a remarkable mAP$_{50}$ of 62.76, which is significantly higher than Point2RBox's 3.41. The poor performance of Point2RBox can be attributed to the large inter-class variability in SKU110K-R, making it difficult to generate suitable patterns. Additionally, PointOBB's SSC loss enforces inter-class consistency constraints, which inherently limits its applicability to single-class datasets like SKU110K-R. The experimental results are provided in the table below.

However, we must also acknowledge that our method performs less effectively on text detection tasks. From our experimental analysis, we observed that the spacing between characters and gaps within individual characters make CPM sample assignment less effective. Even when employing the smallest radius, positive samples inevitably include background regions, reducing detection accuracy. We will include this finding in the limitations section of the paper. Thank you for your professional feedback and constructive suggestions.

Q4. Comparison to TricubeNet.

A4. Thank you for pointing out the resemblance to TricubeNet. We will add a discussion in Section 3.1 of the revised paper to highlight the differences between our CPM and TricubeNet. Specifically:

• TricubeNet uses a symmetric 2D kernel representation, which might offer advantages for precise angle prediction.
• Our CPM, while not symmetric, employs a weakly supervised approach, enabling it to capture approximate object contours with minimal annotation effort.

Dear Reviewers,

May we kindly ask if our responses have addressed your concerns? We look forward to further discussions and feedback from you!

Sincerely,

Authors

Dear Reviewer GcxK,

Could you kindly review the rebuttal thoroughly and let us know whether the authors have adequately addressed the issues raised or if you have any further questions.

Best,

AC of Submission1151

Q1. In related work, it is better to provide a summary of the methods described in section 2.1.

A1. Thank you for your feedback on improving the clarity of Section 2.1. In response, we have revised the related work section to provide a concise summary of existing methods.

Two-stage methods such as Oriented R-CNN, RoI Transformer, and ReDet have introduced mechanisms to handle scale, aspect ratio variations, and rotation explicitly. These methods emphasize alignment features or convolutional adaptations to improve detection accuracy in dense and cluttered environments.

Single-stage methods like R$^3$Det, S$^2$ANet, and DAL focus on improving efficiency without sacrificing precision. Strategies include novel loss functions such as GWD, KLD, and KFIoU, as well as methods for boundary-free angle regression and Gaussian-based loss designs to address discontinuity issues.

In addition to these, RepPoint-based methods predict a set of sample points to describe object extents, providing an alternative to traditional box-based approaches. Methods such as ARC and SASM modify convolutional layers to better handle rotation, demonstrating the evolving strategies for efficient and robust detection.

Q2. How about the effectiveness of the proposed method on other datasets with different object categories and data distributions, such as DIOR?

A2. To further validate the generalizability of our method, we conducted experiments on DIOR-R and SKU110-R. The results, presented in following table, demonstrate that our method consistently improves performance on datasets with diverse object categories and distributions. This indicates the broader applicability and robustness of our approach beyond the DOTA dataset. For more analysis, please refer to Reviewer TgZt's Q2/A2