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.
Limitations
see Weaknesses