This paper presents a simple, self-supervised method for magnifying subtle motions in video: given an input video and a magnification factor, we manipulate the video such that its new optical flow is scaled by the desired amount. To train our model, we propose a loss function that estimates the optical flow of the generated video and penalizes how far if deviates from the given magnification factor. Thus, training involves differentiating through a pretrained optical flow network. Since our model is self-supervised, we can further improve its performance through test-time adaptation, by finetuning it on the input video. It can also be easily extended to magnify the motions of only user-selected objects. Our approach avoids the need for synthetic magnification datasets that have been used to train prior learning-based approaches. Instead, it leverages the existing capabilities of off-the-shelf motion estimators. We demonstrate the effectiveness of our method through evaluations of both visual quality and quantitative metrics on a range of real-world and synthetic videos, and we show our method works for both supervised and unsupervised optical flow methods.
Paper
Similar papers
Peer review
Summary
The paper proposed Lagrangian motion magnification using the pre-trained optical flow network. Magnification loss induces that the optical flow of the magnified frame matches with the optical flow of a given frame by $\alpha$ times; color loss regularizes the color consistency between a given and magnified frames. Test-time adaptation improves the quality of magnification on out-of-domain. Experiments show competitive performance compared to the prior arts in terms of SSIM and the proposed evaluation metrics.
Strengths
1. Simple and effective algorithm to train motion magnification network using the off-the-shelf optical flow network. 2. Given the off-the-shelf optical flow network, this approach enables the training on large-scale unlabeled videos. 3. Targeted magnification and test-time adaptation might provide a better user experience. 4. The proposed method seems to be independent of the architecture of the neural network.
Weaknesses
1. The term should be used carefully. I am not sure that the proposed method can be named "self-supervised" because the off-the-shelf optical flow network, which is used in experiments, is trained by supervised learning. If authors want to use the term "self-supervised", the self-supervised optical flow network should be used in the main experiments and the supervised one would be the strong baseline to be compared; It is not sufficient that the self-supervised optical flow network can be used in theory. 2. Evaluation metric is limited. To justify the underperformance in SSIM, the authors notice this phenomenon in the last sentence in Table 3, "DeepMag explicitly trains for SSIM". The proposed algorithm is also optimized by the proposed evaluation metric, Motion Error. For the same reason, I cannot be convinced about the quantitative results. 3. It is better to include the limitation to use the optical flow network. The optical flow network is inferior to estimate the subtle motion. It is related to the underperformance in the 0.04px subpixel test of Table 3. Thus, the limitation induced by the optical flow network should be investigated as an ablation study because this subtle motion is important in magnification. 4. I think that this proposed method is more general than supervised or self-supervised learning because this is determined by which optical flow network is used. How about using the synthetic data together? 5. Is the network architecture different from DeepMag? As for the control experiments, do the number of parameters affect the performance directly? 6. I think that DeepMag is sufficient as the strong baseline. However, I wonder why Warp Nearest and Bilinear are used, and the more advanced hand-crafted algorithms [A, B] are not used as baselines. It is because evaluation data might contain large motion? [A] Phase-Based Video Motion Processing (SIGGRAPH 2013) [B] Riesz Pyramids for Fast Phase-Based Video Magnification (CVPR 2014)
Questions
My main concerns are the used term and fair comparison. See the weaknesses part. ===== I update my rate from 4 to 6 because the authors will reflect the discussion below.
Rating
6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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
1 poor
Contribution
3 good
Limitations
Limitations and broader impacts are described at the end of the main paper.
Summary
This paper proposes a self-supervised model to solve the Lagrangian motion magnification problem without needing ground-truth labels. The network takes as input the two input frames and a magnification factor that ranges from 1 to 16, and outputs a generated frame that has magnified motion from the first frame. Off-the-shelf optical flow networks are used in loss computation for self-supervision. Test-time adaption has been explored to enhance the generation quality. Experiments show promising results.
Strengths
1. Good writing; overall clear. The studied problem has received arguably less attention in the research community, but the related work section is detailed and well-structured, which especially helps the readers to catch up. 2. The method is very simple and easy to understand. 3. Experiments show promising results. 4. The authors promised to release full code upon acceptance.
Weaknesses
1. The targeting application of this task is not clear. Why is this task important? Which data domains or scenarios do we want it to work? If our goal is just to detect small motions, we can develop optical flow estimation methods that work specifically for small motions. Even for existing state-of-the-art optical flow networks, detecting small motions is generally not a big issue, and it should not be hard to find a way to visualize small optical flow. Why do we need to generate a video? Maybe adding some application examples in the introduction and some results on related datasets will help the reader better understand the background and goal of this task. 2. There are still some confusions on the method. See questions below. 3. Some minor edits. See additional comments below.
Questions
1. As you included the optical flow network in the whole computation graph with gradient computations, did you freeze its weights to make sure it does not shift away? If so, please state so explicitly in the paper. 2. Line 185-186: Why is the positional encoding conditioned on the factor $\alpha$? How do you do it? It does not make sense to me if no explanations are given. 3. Test-time adaptation: this trick adds in overhead on inference time. How efficient is it? 4. Did you tackle the occlusion issue? I believe this is an issue both for new frame generation and optical flow estimation. A simple warping using optical flow may be good enough if there are no occlusions at all. Additional comments: 1. Line 26-28: "The datasets and learning procedures that are used by these models are designed to be general-purpose, with a particular focus on ensuring that they apply to a variety of motions, objects, and scenes". Your cited methods are all supervised methods, which are usually trained on large synthetic datasets that could be totally different from real use cases. The generalization ability of these models is still an open question. I think a better idea is to weaken this statement. 2. Line 32-34: "And generate a new image pair whose predicted optical flow is $\alpha$ times as large as that of the input". Maybe add that the new image pair should share the same reference frame from the input, to avoid confusions. 3. Line 33-34: add $(\alpha \geq 1)$ to be more clear. Your method also makes sense even if $\alpha < 1$, so it is better to clarify your work cases. 4. Line 38: "it" -> "our method". 5. Line 61: need a citation for "Eulerian approaches". 6. Line 69: "it uses it" -> "we use it". 7. Line 115: It is better to make the symbol $x$ bold like $\mathbf x$ since it is a vector. 8. Eq 1: Maybe add that this equation assumes no occlusions. 9. Fig 6 caption: "subset" -> "subsets". 10. Repeated references: [21] and [22], [37] and [38].
Rating
5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.
Confidence
3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
Soundness
3 good
Presentation
3 good
Contribution
2 fair
Limitations
Maybe need to add occlusions as a key limitation
Summary
The paper introduces an optical-flow-based lagrangian motion magnification method, learned through self-supervised learning. The architecture is very simple -- just a U-Net that inputs two temporally consecutive frames and outputs a motion-magnified image. To train the U-Net, the method uses an off-the-shelf optical flow method, estimates the motion between two frames, and considers the motion as real motion of the scene. Then it penalizes the difference between the estimated magnified motion (constant * estimated motion) and the motion between the reference image and motion-magnified image (i.e. output image). The method demonstrates both good quantitative and qualitative results.
Strengths
+ Comprehensive related work The paper provides a comprehensive literature survey, which helps understand previous related work and where this paper positions among them. + Implementation details Sec. 4.1, Sec. 4.2, and supplementary material provide sufficient implementation details so that it's easy to understand the choices of hyper-parameters, training configuration, dataset curations, and training details. Parts of source codes are also included in the supplementary material, which all help reproduce the proposed methods. + In-depth evaluation Fig. 5, Fig. 6, Table 2, and Table 3 provide in-depth evaluation of the proposed method and related methods on real-world and synthetic videos. Given that there is no public benchmark and the difficulty of evaluation on this topic, the paper tries its best on providing sufficient evaluation.
Weaknesses
- How to handle occlusion and disocclusion? It seems already stated in the limitation section, but I wonder if the method doesn't explicitly handle occlusion and disocclusion. If it doesn't, then can it be a problem? Does U-Net learn to handle them to some extent? When watching the supplementary video, the model doesn't seem to output hallucinated appearance around the disoccluded region, which seems good. - Worse SSIM in Table 3 In Table 3, compared to DeepMag, all metrics are better except for SSIM. I wonder why it's the case. What makes the DeepMag's SSIM better than the proposed method? - Moving background? In the supplementary video (1m14s and 2m12s), I am wondering why the background has motion and it's moving? Is it due to that the optical flow method hallucinates motion in the background and it's used during the training? Can this problem be resolved without using the target segmentation mask? By the way, this is another question: what if $L_{mag}$ in Eq. (3) is applied to the target segmentation objects only and penalizes the background motion to be zero? Can it produce better results and prevent the background from moving? There are some unresolved concerns but the strength outweighs the weaknesses for now. I would like to give Borderline Accept for now, but the rating could change after the discussion phase. --- All concerns are resolved. Thus I am updating my rating to 7. Accept.
Questions
- Usage of the discriminator loss? I wonder if using an adversarial loss from GAN would help more realistic appearance where flow is not accurate, occlusion happens, or artifacts occur. - I wonder if the collected data or at least the information on train/test splits will be available if the paper is accepted.
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
- Probably another limitation would be that the method's success depends on the off-the-shelf optical flow and segmentation methods.
Summary
This paper uses the classical method of Lagrangian to self-supervise the task of motion magnification. Thanks to the proposed self-supervision-based technique, the proposed method can also be adapted during the test time. As shown in Figure.1, the proposed method is simple, where the optical flow vectors of videos before and after magnification are compared. The optical flow of the motion amplified video is compared to the scaled (by the amplification factor) optical flow of the original video, to derive the magnification loss. To make the output video color consistent, color loss is also used. Videos are provided in the supplementary material for qualitative analysis.
Strengths
1. The method presented in this paper is simple, straightforward, and meaningful. 2. The experimental evaluations validate the proposed method. Supplementary videos are helpful. 3. The source code is also provided in the supplementary, which further highlights the simplicity. 4. Limitations of the method are well discussed, and failure cases are shown. 5. The paper is well-written and easy to follow.
Weaknesses
1. The proposed method largely depends on the pre-trained optical flow network. 2. Given the nature of the addressed problem, its evaluation is known to be difficult. This is reflected in the experiments. 3. The experiments are conducted in a relatively small amount of video frames, and the paper discusses “out of the distribution” and “test time adaptation”. It would be interesting to see how the method generalizes when training on a large number of videos, before proceeding to discuss the rest.
Questions
How does the proposed method behave with low-quality optical flow? Does the method improve its performance in difficult videos, when trained on a larger amount of data? I can imagine that when trained on a large collection of videos, the method may generalize and hence also perform well in some difficult cases.
Rating
6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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
4 excellent
Contribution
3 good
Limitations
The authors adequately addressed the limitations.
Summary
This paper focuses on learning a pair-wise motion magnification model in a self-supervised manner. The authors employ recent optical flow models to estimate the flow fields between the original and the motion magnified image pairs. The UNet concatenates a sinusoidal encoded magnification factor with the original images and generates the magnified image. The learning process of the UNet is facilitated by a loss function that enforces consistency between the original and magnified flow fields, as well as the consistency between the backward warped images. To demonstrate the effectiveness of the proposed method, the authors curate a large-scale real-world training set. They conduct evaluations both quantitatively on a synthetic dataset and qualitatively on real-world data. The results demonstrate superior performance over previous supervised methods learned on synthetic data.
Strengths
- The paper is well organized and written. Sec.1 introduces the problem effectively and motivates the design of the method. Sec.2 provides a brief but comprehensive review of the previous methods. In addition to the overall organization, sufficient details (including the code) are given for a better understanding of the method, such as the footnote on page 5. - The method itself is simple and effective: - A simple UNet is a compact solution that avoids complicated operations, e.g., explicit optical flow estimation and inpainting. - The magnification factor is concatenated with the input image pair after sinusoidal encoding, which enables regionally varying magnification. This is difficult for a single magnification factor as in [33]. - The self-supervised learning losses enable online adaptation to a specific sequence for better quality. - The evaluation is comprehensive and achieves significant improvement over previous methods. - The curated large-scale real-world dataset will encourage further research in this topic.
Weaknesses
I do not see significant weakness of the paper since it is simple and effective. The only missing piece I come up with is that since the model itself is simple, the author could make some deeper analysis of a learned UNet to understand the underlying mechanism of the model.
Questions
- In lines 83-84, the authors claim that the method's capabilities are more similar to those of Lagrangian methods. However, the internal workings of the model remain unknown. It would be interesting if the authors could conduct some analysis, such as investigating whether some UNet layers implicitly contain motion cues. - The warping loss compares warped images. What would happen if we additionally, or only, compare two warped images with I_0? - In the supplementary video, is the magnification factor applied only to the ground in the jumping sequence?
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
4 excellent
Presentation
4 excellent
Contribution
3 good
Limitations
The limitations have been addressed adequately in the paper.
Thanks for the response! I still have questions on the following topics. **Importance of Motion Magnification** If the goal of the task is to detect and highlight small motions in different applications, we could just visualize the magnitude of the optical flow field. Ideally, most part of the visualization will be totally black due to zero motion, and only the small motion part is highlighted. This should be obvious enough if the goal is just to detect small motions, so why do we need to generate a new video with magnified motion? For medical workers or engineers, this type of visualization should be easier to read than original color images. In addition, generating a new video with magnified motion will most likely create new occlusions, so a part of the information in the original input may be lost. For example, what if the magnified motion covers a part of the background that also contains another small motion that needs to be detected? In comparison, a simple visualization of per-pixel motion such as optical flow should be able to highlight every small motion in the image. I suggest that the authors could use some examples from targeting applications as demo, instead of the current "baby" and "cats" examples (looking like naive toy cases). That will help better explain the background applications of the task. **Occlusion and Dis-occluson** Based on my experiences in optical flow estimation, occlusion is still a very challenging issue and also a major source of error for most of the latest optical flow models including RAFT. The correspondences of occluded pixels are not perceivable in the second frame, so the model can only "guess" the correct flow based on smoothness and other cues. I think the main reason that occlusions do not hurt too much in your case is that the occlusion regions are also very small if the motions are generally small, so masking the occlusion region does not make big differences. Maybe you could visualize your occlusion masks and argue in this direction. Anyway, stating an optical flow model can handle occlusions well could be a bit dangerous. Still, the paper should acknowledge the issues on occlusions even if you could argue it does not hurt your specific task too much. Many of the equations and losses do not work at occlusion regions, so it will be confusing if you do not mention occlusions there at all.
**Importance of Motion Magnification** We agree with you that if our goal were to detect or localize small motions we could just plot an optical flow field. The goal of motion magnification is instead to visualize these small motions. For very simple motions, this may not be necessary and your proposal would work well, but often people need to understand complex motions that can’t be visualized well from optical flow alone. For example, in Figure 2c of our paper we motion magnify the “pole” clip, originally from [1]. Looking at the patch visualization it is clear that the column is vibrating, but more precisely motion magnification shows that it is vibrating at *two different modes*. One lower frequency mode at a higher amplitude, and one higher frequency mode at a lower amplitude. The superposition of the two modes produces the distinct “squiggly sine wave” patch visualization. Understanding this from just optical flow visualization alone would be quite difficult, but magnifying motions makes this abundantly clear. Other examples of motion magnification visualization can be found in the cited PNAS article [2], where the authors use and validate the technique for a number of scientific applications. This includes visualization of the modal shapes of a pipe (Fig. S3) and a lift bridge (Fig. 3), vibrations in ear tissue (Fig. 2), and deformations in a metamaterial under forcing (Fig. 4). In addition, we point out that at this year’s SIGGRAPH conference the original Eulerian motion magnification paper [3] was awarded the “Test-of-Time Award,” highlighting the impact of this visualization technique. Finally, with respect to your suggestion of replacing the “baby” sequence in our paper, we note that it is a "classic" example used in prior work [3,4,5,6]. We will edit our draft to better explain the applications of motion magnification and include more practical applications of the technique, and we thank you for your suggestions. **Occlusion and Disocclusion** We also agree with you that occlusions and disocclusions are a challenge in the context of motion magnification methods. Our discussion above was not meant to imply that we’ve solved the problem of occlusions, but rather to point out our method does not ignore occlusions entirely and is able to reason about them through the optical flow model. As you correctly point out, these models are not perfect at predicting flow in occluded areas and therefore large occlusions may be challenging. However, because our method is agnostic to the form of tracking used we believe it can benefit from future progress in optical flow estimation. Additionally, empirical results demonstrate that motion is amplified well despite these concerns. This is likely due to your point that motions we consider are small. Overall, we thank you for your helpful thoughts on this problem and your insightful recommendations. We will revise our draft to include a thorough discussion of these points. Again, thank you for your suggestions and please let us know if you have any remaining questions or would like to discuss a point further! [1] “Structural Modal Identification through High Speed Camera Video: Motion Magnification.” Justin G. Chen, Neal Wadhwa, Young-Jin Cha, Frédo Durand, William T. Freeman, Oral Buyukozturk. *Proceedings of the 32nd International Modal Analysis Conference (2014).* http://people.csail.mit.edu/mrub/vidmag/papers/Chen_Imac_2014.pdf [2] “Motion microscopy for visualizing and quantifying small motions.” Neal Wadhwa, Justin G. Chen, Jonathan B. Sellon, Donglai Wei, Michael Rubinstein, Roozbeh Ghaffari, Dennis M. Freeman, Oral Büyüköztürk, Pai Wang, Sijie Sun, Sung Hoon Kang, Katia Bertoldi, Frédo Durand, and William T. Freeman. *Proc. Natl. Acad. Sci., 114 (44) (2017), pp. 11639-11644.* https://www.pnas.org/doi/full/10.1073/pnas.1703715114 [3] “Eulerian Video Magnification for Revealing Subtle Changes in the World.” Hao-Yu Wu, Michael Rubinstein, Eugene Shih, John Guttag, Frédo Durand, William T. Freeman. *ACM Transactions on Graphics, Volume 31, Number 4 (Proc. SIGGRAPH), 2012*. [4] “Phase-based Video Motion Processing.” Neal Wadhwa, Michael Rubinstein, Frédo Durand, William T. Freeman. *ACM Transactions on Graphics, Volume 32, Number 4 (Proc. SIGGRAPH), 2013.* [5] “Riesz Pyramids for Fast Phase-Based Video Magnification.” Neal Wadhwa, Michael Rubinstein, Frédo Durand, William T. Freeman. *IEEE International Conference on Computational Photography (ICCP), 2014.* [6] “Learning-based Video Motion Magnification.” Tae-Hyun Oh*, Ronnachai Jaroensri*, Changil Kim, Mohamed Elgharib, Frédo Durand, William T. Freeman, Wojciech Matusik. *European Conference on Computer Vision (ECCV), 2018.*
Thanks for your detailed response! I now have a better understanding about the background of this task. I think this paper could be more accessible for more specialized conferences such as SIGGRAPH or CVPR. For NeurIPS audiences, it is better to include those technical backgrounds so that the readers could understand the goals and expectations of this task.
Thank you so much for your reply and helpful suggestions to help us refine the paper. We will include relevant technical background in our manuscript and tailor it to the NeurIPS audience. If you have any additional suggestions to improve our paper please don't hesitate to share.
Response of Rebuttal by Authors
Thank the authors for responding to the comments. **The Term "Self-Supervised"** I am not sure and convinced about the authors' opinion. For meaningful discussion, I want to hear the author's opinion about the below example. Let's think about weakly-supervised semantic segmentation [A]. I think that this is similar setting of this paper. - Train Model A on task A by supervised learning - This paper: Optical flow - Weakly-supervised semantic segmentation: Classification - Apply Model A on ML algorithm B not using the ground truth of Task B - This paper: Motion Magnification - Weakly-supervised semantic segmentation: Segmentation In my opinion, "Weakly-Supervised Learning" is more appropriate to this paper because this work exploits more cheap ground truth of optical flow rather than motion magnification. Can authors provide their opinion about this? My big concern is the use of terminology in this work. If this is tackled, I will raise my score. And where can I find the result of the method trained by ARFlow in Tables 2 and 3? I would appreciate for the author to provide the line number. [A] Evaluation for Weakly Supervised Object Localization: Protocol, Metrics, and Datasets, TPAMI 2022 **Evaluation Metrics** **(i)** I understand the difficulty of evaluating motion magnification, like the evaluation of a generative model. Since there exists no public benchmark of motion magnification, it might be difficult to request the construction of real-world dataset; move the object by x and alpha x exactly and make data pair. I think that "We do better on motion error, but lag slightly behind on SSIM, which DeepMag explicitly trains for" should be removed for the following reasons: (1) Question about the superior performance on motion error because the authors' method is trained for this, (2) Doubt of the validity of the evaluation metric from (1). I believe that it does not degrade authors' work. **(ii) Joint Supervised and Self-Supervised Training** This question is followed by the performance number of motion errors and SSIM. It is because joint training will improve both metrics together. I agree with authors' response. **Architectural Differences** I think that reporting the number of parameters is sufficient though the experiment is the direct way. **Additional paper** I suggest this paper: "Video Motion Magnification to Improve the Accuracy of Vision-Based Vibration Measurements". It is because this paper can support the need for motion magnification, and the learning-based method would be superior compared to other methods.
Response of Rebuttal by Authors
I checked Table 2 and Table 3 from the attached pdf. In my opinion, if the results are replaced by ARFlow, I respect the self-supervised method. In addition, the results in the main paper can be the strong baseline or upper bound. I believe that the number might not be important compared to DeepMag in this case, and the approach is cool. This method is more general, so the trained method, such as supervised, self-supervised, and weakly supervised does not degrade this idea. However, if the definition or terminology is arguable, I think that It should be discussed carefully. Can authors provide a discussion about this point?
Thank you for reading and responding to our comments! **The Term “Self-Supervised”** In light of the unsupervised ARFlow results would the following changes to the manuscript be satisfactory? - Include fully trained ARFlow results under the label of fully self-supervised - Keep the RAFT results, but refer to them as “weakly supervised.” In addition indicate that they serve as a rough upper bound for a fully self-supervised method. - Include a careful discussion of self-supervision and weak supervision in the context of our proposed method We apologize for any confusion and please let us know what you think. **Evaluation Metrics** We agree with your assessment that Motion Error is not necessarily the better metric and SSIM is not necessarily deficient. Both metrics have different trade-offs and are useful in different scenarios. We can remove the phrase you mention and include an in-depth discussion of the metrics that we use. **Architectural Differences** We will provide more details on the architectures used in our paper, including the number of parameters. **Additional References** Thank you for the paper reference! We will include this in our bibliography and discuss it in our related work section. Again, thank you for your comments. Please let us know your thoughts on the above proposed edits to the manuscript. We would be happy to discuss any other points you may have!
I appreciate the authors' response. If the authors clear the confusion as the authors mentioned above, it would resolve my concerns, and I will update my rating.
Thank you so much for your feedback! We note that the NeurIPS 2023 submission policies do not allow us to update our manuscript during the rebuttal process, but we will certainly make the changes noted above and clear any confusion with respect to supervision in our manuscript. We thank you for your time and thoughtful feedback, and your willingness to update your rating.
Thanks for the detailed responses. It resolves my concerns. I will also read other reviews and update my rating accordingly after the discussion period!
Dear Authors: Thank you for the response and all my concerns have been addressed.
Discussion follow-up
Thank you for your reply. I suggest to include the experiments with different quality of optical flow. I think, these are simple experiments to conduct, therefore should be addressable in the camera-ready. Hence, I keep my review towards accepting this paper.
Decision
Accept (poster)