H2RBox-v2: Incorporating Symmetry for Boosting Horizontal Box Supervised Oriented Object Detection

With the rapidly increasing demand for oriented object detection, e.g. in autonomous driving and remote sensing, the recently proposed paradigm involving weakly-supervised detector H2RBox for learning rotated box (RBox) from the more readily-available horizontal box (HBox) has shown promise. This paper presents H2RBox-v2, to further bridge the gap between HBox-supervised and RBox-supervised oriented object detection. Specifically, we propose to leverage the reflection symmetry via flip and rotate consistencies, using a weakly-supervised network branch similar to H2RBox, together with a novel self-supervised branch that learns orientations from the symmetry inherent in visual objects. The detector is further stabilized and enhanced by practical techniques to cope with peripheral issues e.g. angular periodicity. To our best knowledge, H2RBox-v2 is the first symmetry-aware self-supervised paradigm for oriented object detection. In particular, our method shows less susceptibility to low-quality annotation and insufficient training data compared to H2RBox. Specifically, H2RBox-v2 achieves very close performance to a rotation annotation trained counterpart -- Rotated FCOS: 1) DOTA-v1.0/1.5/2.0: 72.31%/64.76%/50.33% vs. 72.44%/64.53%/51.77%; 2) HRSC: 89.66% vs. 88.99%; 3) FAIR1M: 42.27% vs. 41.25%.

Paper

Similar papers

Peer review

Reviewer Pshp7/10 · confidence 4/52023-06-13

Summary

This paper proposes to exploit the reflection symmetry as a new supervision to HBox-supervised oriented object detection. Several modifications are made to adapt the new self-supervised (SS) branch, including removing the angle subnet in the weakly-supervised (WS) branch and a CircumIoU loss for box regression. Experiments show a clear state-of-the-art result to previous Hbox-supervised oriented object detectors.

Strengths

1. The paper tries to exploit the reflection symmetry to improve the Hbox-supervised object detector, which is interesting. 2. The paper is well-motivated and easy to understand. 3. The proposed H2Rbox-v2 achieves state-of-the-art performance in the Hbox-supervised oriented object detectors.

Weaknesses

Major concern: 1. The soundness of using reflection symmetry is somewhat low. According to page 5 line 130, if the network $f_{nn}(\cdot)$ subjects to both flip and rotate consistencies, then $f_{nn}(I_0)=\theta_{sym}$. I don't know why this equation holds. Clearly, we don't know the symmetrical axis of the image $I_0$. And if we rotate the image by a random angle $\theta$ in the Way2 of Fig.3, then $f_{nn}(I_2)$ would never be equal to $\theta_{sym}$. Even if the symmetrical axis is known, this reflection symmetry learning can only theoretically and ideally work under the setting of a single symmetric object. As a matter of fact, it would be more rational that the converse proposition holds - "If the network $f_{nn}(\cdot)$ can always predict the symmetrical axis $\theta_{sym}$ of the image, then, it must subject to both flip and rotate consistencies." I suggest the authors reconsider the logic here. 2. Due to weakness 1, the novelty is also limited since one can easily conclude that H2Rbox-v2 is to add one more additional branch with a new view generation upon H2Rbox, i.e., the left column of Fig. 2(b) ("vertical flipping"). The right column is random rotation which is similar to the SS branch of H2Rbox. This is what I believe the model brings us performance gains, but not the so-called reflection symmetry learning. The experiments cannot support that the reflection symmetry is important, but can only verify the effectiveness of adding the vertical flipping branch, which is supported by Tab. 6. It is still unclear for the relationship between reflection symmetry learning and multi-branch self-supervised learning. 3. The ablation studies provide a weak explanation for the effectiveness of the proposed components. In Tab. 4, adding the $l_s$ loss only produces near zero mAP on DOTA, while it is much higher on HRSC. Why did such drastic fluctuations occur? The PSC coder is just an angle encoder-decoder module, which is originally proposed to improve the performance of oriented object detectors. It would not be a critical factor affecting the performance, as demonstrated by the PSC paper, ~2 mAP at most. This paper shows PSC is crucial in the proposed method. That is to say, without PSC, the proposed method produces much poor performance. This largely weakens the soundness of the designs in the SS branch. Additionally, PSC imposes the loss on two encoded phase-shifting patterns of angles, which inherently solves the boundary discontinuity, while the proposed SS branch decodes the phase-shifting patterns to angle and then calculates the loss on two angles, which introduces the boundary discontinuity again, thus undermining the contribution of the snap loss. 4. It is confusing to me that Fig.2 shows the angle is predicted by the SS branch, while all the other properties are predicted by the WS branch. That is to say, the inference time of H2Rbox-v2 (WS+SS) should be double of the WS branch, while H2Rbox only needs the WS branch for inference. I'm concerned about why the FPS of H2Rbox-v2 is equal to H2Rbox as shown in Tab. 2. The authors need to describe the inference process clearly. Minor concern: The baseline model is unclear in the ablation studies. In Tab. 4, if the baseline model is "w/o PSC and $l_s$", then what is the angle coder used in the baseline? And what are the angles in calculating Smooth-L1 loss? The same question emerges in Tab. 5 and Tab. 6. Is that mean all the baseline models of the tables adopt the optimal strategies of the other tables? If so, then the best results of Tab.6 seem to be not in line with the others. Overall: While the proposed method achieves state-of-the-art performance, my initial rating score is 4. I have some concerns about the soundness of the reflection symmetry learning and the ablation studies. It is of vital importance to give an explanation and analysis of the rationality of the proposed method.

Questions

Suggestions: 1. For the weaknesses 1 and 2, I suggest the author to rethink the logic of the analysis of Sec. 3.1. It would be better if we could restate the analysis in page 5 line 130 to the converse proposition - "If the network $f_{nn}(\cdot)$ can always predict the symmetrical axis $\theta_{sym}$ of the image, then, it must subject to both flip and rotate consistencies." And then the contrapositive - "if the network subjects to neither flip consistency nor rotate consistency, it cannot predict the symmetrical axis of the image." Therefore, we need to enhance both flip consistency and rotate consistency. H2Rbox considers the rotate consistency only, which leads to sub-optimal results. 2. There is a missing experiment, i.e., $\lambda=0$ in Tab. 6.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

see Weaknesses

Reviewer tfHb5/10 · confidence 5/52023-06-15

Summary

This paper proposes an advance solution for using horizontal bounding boxes as supervision to learn oriented object detectors. The proposed method H2RBox-V2, which is a modification of the recent work H2RBox, has some technique novelty and contribution as it jointly uses the weakly- and self-supervised branches to learn the rotation angle. Experiments are validated on DOTA, HRSC and FAIR1M, and show state-of-the-art performance.

Strengths

+ The overall idea to learn the rotated angles from both weakly- and self- supervised branch is very interesting. + The task itself, to learn rotated bounding boxes from horizontal bounding boxes, is a new setting for the rotated object detection community. + The performance improvement against prior arts is significant.

Weaknesses

- Unclear methodology design and presentation. (1) The authors claim to propose a new CircumIoU loss for this framework. Unfortunately, in the methodology section, especially Sec3.4, there is no term ‘CircumIoU loss’, and it is impossible to know which loss is the claimed novelty. (2) Besides, given this unclear description, it is very difficult to know its difference or novelty against some prior works such as [a], which also optimizes the detector from rotation angles. [a] Arbitrary-Oriented Object Detection with Circular Smooth Label. ECCV 2020. Other minor issues and comments for improvement: - Please provide some rotation angle distribution visualizations from the proposed framework. - Fig.1, column2 and column4. The predictions from the proposed method and SAM-RBox seem not to have much difference. Please consider use more representative figures. - Table 1, the performance gap. Please explictly mention which dataset it is that leads to 3.41% drop and 0.07% improvement. - Fig.2 looks very crowded. It can be better polished, and the size of each content can be made more fit. - For loss function, please use \mathcal{} to distinguish it from scalars.

Questions

The authors have addressed my questions in the rebuttal well. ______ before rebuttal _____________ Q1: Issues on the flaw of the technique framework. Both experimentally and theoretically. Q2: The details, design and clear presentation of the CircumIoU loss. Q3: systematically address the unfair experimental comparison issue.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The authors have addressed my concerns well. ______________________before rebuttal ___________ The limitations are not properly discussed. The reviewer believes the technique flaws as mentioned in weakness part is more critical to the proposed framework.

Reviewer W3Ay7/10 · confidence 5/52023-07-04

Summary

This paper introduces H2RBox-v2, an innovative approach to further bridge the gap between HBox-supervised and RBox-supervised oriented object detection. It seeks to address the limitations of the original H2RBox model, which required high-quality annotations and large training datasets, and was incompatible with rotation augmentation. H2RBox-v2 augments the original model by adding a self-supervised branch that learns object orientations from inherent visual symmetry, and a weakly-supervised branch that incorporates a new CircumIoU loss to allow for random rotation augmentation.

Strengths

- **Originality:** While the paper builds upon existing work (specifically H2RBox), it infuses new concepts, the most significant being the utilization of symmetry for angle regression. This concept is highly innovative and introduces a novel angle of approach for oriented object detection. This creativity in applying the natural property of symmetry to enhance detection accuracy sets this work apart from previous methods and broadens the boundary of the field. - **Quality:** The proposed H2RBox-v2 exhibits enhanced performance on various datasets compared to its predecessor, extensive ablation experiments have demonstrated the importance of each module. - **Clarity:** The paper is well-structured and clearly presents the methodology and results, making the contributions of this research easily understandable. - **Significance:** Bridging the gap between HBox-supervised and RBox-supervised oriented object detection significantly reduce the annotation costs. - H2RBox-v2 has shown improved performance on various datasets and is specifically designed to cope with situations where H2RBox-v1 may underperform. This makes it a valuable contribution to real-world applications of oriented object detection.

Weaknesses

- **Theory is not fully comprehensive:** *L138-142* only discussed the case where a single instance is contained in an image. However, in reality, remote sensing images can contain many objects, especially in dense scenarios. Despite the theoretical proof being not fully comprehensive, their simplicity of idea and empirical effectiveness seem to be sufficient according to me. - **Training overhead not detailed:** It is crucial that the model does not introduce additional overhead during the testing phase, but the paper does not discuss the computational overhead and the time taken for the model training process in detail

Questions

- Would it be more effective to conduct the comparative experiments on the DOTA dataset, regarding the assertion that H2Rbox-v1 requires more training data than v2? Specifically, by training both v1 and v2 with varying percentages of DOTA data (namely 10%, 20%, and 50%), we could gain a more robust comparison of their respective performances. - When the object is not in the center of the image, are Way1 and Way2 described in L124-132 still equivalent? - In remote sensing scenarios, some objects may not possess symmetry, such as swimming pools and harbors. What impact does this have on the theory?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Nothing to report.

Reviewer k1cY7/10 · confidence 5/52023-07-05

Summary

This paper proposes a new horizontal box-supervised rotation object detection detector. The proposed detector consists of two modules: a self-supervised regression branch for angle regression and a weakly supervised branch for horizontal box regression. This method is more simpler and shows clear improvement over the previous one.

Strengths

- The paper fully considers the independence and correlation between angle and horizontal box in rotation object detection. It only uses angle regression in self-supervised regression while using Circumscribed RBox IoU to associate angle regression and horizontal box regression in weakly supervised regression to obtain the rotated box. - The paper greatly improves the performance of the hbox-supervised detector, approaching or even reaching the performance of some rbox-supervised detectors. - The paper is well written, and the method is clear and well described.

Weaknesses

- Figure 2 seems a bit complicated, it would be better to highlight the main points to make it clearer. - Although this article has achieved impressive results, it is important to consider the robustness of the model when horizontal bounding box annotations are not accurate enough, the article's experiments do not explicitly demonstrate the robustness towards inaccuracies in horizontal bounding box annotations.

Questions

- I have a question about the self-supervised branch in the proposed method. Why was it designed with two perspectives instead of using one perspective and randomly selecting multiple variations under that perspective? Additionally, would using a more perspective, such as different rotation angles, combinations of rotation and symmetries, or scaling variations, lead to performance improvements? - Regarding Table 4, I noticed that the experiments on DOTA and HRSC datasets show inconsistent results. Specifically, in the Dota dataset, not using PSC will result in an incorrect result, while in the HRSC dataset, not using PSC only results in a performance loss. This raises the question of whether PSC only improves performance or if it has an indispensable impact on the model.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

N/A

Reviewer tfHb2023-08-10

Response to author rebuttal

Many thanks for the time and effort from the authors. Most of my concerns have been properly addressed, and I realized that the orignal rating before rebuttal is down-graded. Especially, the clarification of significant improvement against H2RBox under *the same and fair setting* makes this work much better than the previous grade. However, before the discussion period ends, still some minor concerns regarding **Q1.b** : - Indeed the authors have clarified that both backbones are weight-sharing before angle prediction. Does it involve any additional layers or parameters to map the frozen backbone feature to the angle? - If the answer of first question is true, then, empercially extract the frozen features and learns the mapping between angle and backbone through the other branch may have some unforseen impact. - In this regard, it would be better for the authors to show some empercial outcomes on this aspect, so that my last remaining concern is resolved. I am willing to improve my rating to the accept threshold if these final concerns can be properly addressed in the following week.

Authorsrebuttal2023-08-11

Thanks for your prompt and responsible response. We try to get your point and our tentative response is as follows. **Q1: Does it involve any additional layers or parameters to map the frozen backbone feature to the angle?** Following backbone, there is an angle head, consisting of several convolution layers and a PSC decoder, to map the feature to the angle (see Fig.2b). The angle heads in different branches also share the same parameters. In fact, all heads, including regression, classification, center-ness and angle, their parameters are also shared, which is consistent with the classic detectors, i.e. RetinaNet and FCOS. **Q2: Extract the frozen features and learn the mapping between angle and backbone through the other branch may have some unforeseen impact** Please note that our backbone is not frozen **during training** -- the backbone, angle head, and other heads (i.e. regression, classification, center-ness) are updated together by SS and WS losses. Once the training is complete, both the backbone and the heads are frozen, and there is no need to generate SS views or learn the angle mapping **during inference**. In oriented object detection, it is common to predict angle in a separate head/branch, e.g. in your mentioned prior work CSL, the angle is also separately predicted by an angle head which is trained by an independent angle loss. The detection results and the performance show that our design works well. Please let us know if we misunderstood your questions, and also let us know if there are further questions.

Reviewer tfHb2023-08-11

Response to author rebuttal (Round 2)

Many thanks for the authors to provide informative and timely response. - Regarding **Q1**, as all the parameters are all shared, then my concerns have been well addressed. - Regrading **Q2**, yes, I think the reviewers have clearly catched my point. Although there is no emperically justification, I think the explanation is responable. After the rebuttal and the clarification, I think now I hold more positive than negative view on this work. I will revise my comments and raise my score in the following day, as it takes some time. Meanwhile, I would kindly suggest the authors: - if possible, please revise Fig.2 in the camera-ready version in a more clear way. - if possible, CircumIoU loss part can be presented in a more informative way in the camera-ready version. Finally, I would like to thank the authors again for their patient and clear response.

Authorsrebuttal2023-08-13

We will carefully revise the paper and fix these issues. Thanks again for your recognition and valuable suggestions on this paper!

Reviewer k1cY2023-08-13

My final decision

Thanks for the response. I think it all makes sense, and glad to see the authors added the experiment of using one perspective/branch, leading to an improvement in H2Rbox-v2’s performance compared to v1 without any adverse effects. Thanks for the good paper. Now I do not find any other fatal problems, and I will increase my rating from 6 to 7.

Authorsrebuttal2023-08-13

We will carefully revise and prepare the final version. Thanks again for your nice words and constructive suggestions on this paper!

Reviewer Pshp2023-08-13

Response to authors

Thanks for the detailed rebuttal. Most of my concerns are addressed. The main paper should include the experiment of Q2.a. The authors should add analyses on Q3.a, Q3.B, and Q3.c, and fix the issue of Q5.b. I'm willing to improve my rating score from 4 to 7.

Authorsrebuttal2023-08-13

We will carefully fix these issues in preparing the final version. Thanks again for your nice suggestions and efforts on this paper!

Reviewer W3Ay2023-08-16

I would thank the authors for addressing my concerns. Given my current rating of 7, I intend to maintain it, unless other reviewers introduce new issues that warrant reconsideration.

Authorsrebuttal2023-08-16

We will carefully prepare the final version. Thanks again for your recognition and valuable suggestions!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC