Unsupervised Optical Flow Estimation with Dynamic Timing Representation for Spike Camera

Efficiently selecting an appropriate spike stream data length to extract precise information is the key to the spike vision tasks. To address this issue, we propose a dynamic timing representation for spike streams. Based on multi-layers architecture, it applies dilated convolutions on temporal dimension to extract features on multi-temporal scales with few parameters. And we design layer attention to dynamically fuse these features. Moreover, we propose an unsupervised learning method for optical flow estimation in a spike-based manner to break the dependence on labeled data. In addition, to verify the robustness, we also build a spike-based synthetic validation dataset for extreme scenarios in autonomous driving, denoted as SSES dataset. It consists of various corner cases. Experiments show that our method can predict optical flow from spike streams in different high-speed scenes, including real scenes. For instance, our method gets $15\%$ and $19\%$ error reduction from the best spike-based work, SCFlow, in $\Delta t=10$ and $\Delta t=20$ respectively which are the same settings as the previous works.

Paper

References (49)

Scroll for more · 37 remaining

Similar papers

Peer review

Reviewer PCU56/10 · confidence 5/52023-06-10

Summary

The paper titled "Unsupervised Optical Flow Estimation with Dynamic Timing Representation for Spike Camera" presents a novel approach for unsupervised optical flow estimation using spike cameras. The authors propose a method that leverages dynamic timing representation to estimate optical flow without the need for explicit supervision. They demonstrate the effectiveness of their approach through extensive experiments and comparisons with existing methods.

Strengths

1. The paper introduces a unique and innovative approach to unsupervised optical flow estimation. By utilizing spike cameras and dynamic timing representation, the authors provide a fresh perspective on tackling this problem. This can potentially open new avenues for research and development in the field. 2. The paper presents the methodology in a clear and concise manner. The authors provide sufficient details and explanations, making it easier for readers to understand the technical aspects of their approach. They also include relevant diagrams and figures to aid in comprehension.

Weaknesses

1. The paper lacks a thorough discussion of the limitations of the proposed method. While the authors demonstrate its effectiveness through experiments, it would be beneficial to address potential challenges or scenarios where the method might not perform optimally. This would provide a more comprehensive understanding of the approach. 2. The authors primarily focus on benchmark datasets and quantitative evaluations, but they do not provide real-world examples or scenarios where their method could be applied. Including such examples would enhance the practical relevance of the research and highlight its potential impact.

Questions

1.How does the proposed method handle occlusions or complex scenes where multiple objects are moving simultaneously? Are there any specific limitations or challenges in such scenarios? 2. How does the dynamic timing representation used in the spike camera contribute to the accuracy of optical flow estimation compared to traditional frame-based approaches?

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

3 good

Contribution

3 good

Limitations

The proposed method's performance and generalizability might be limited to specific datasets or conditions. It would be valuable to explore its effectiveness on a wider range of datasets, including more challenging and diverse scenarios.

Reviewer Hpzy6/10 · confidence 4/52023-06-28

Summary

The paper proposes a method for estimating optical flow from event streams captured with a spike camera. The approach is based on unsupervised learning and achieves good results in the comparison to other works in the SotA, including conventional and event-based approaches. Authors collect their own dataset and evaluate on it, specially, in an attempt to show the advantage of finely estimating for low and high-speed regions (although, this is not clearly demonstrated with the current experimental section). Finally, authors also include a section with an ablation study to assess the impact of the different parts of the model (loss terms) and the supervised vs. unsupervised learning approaches.

Strengths

+ Authors prepare and collect a dataset with is always a hard but valuable task + The unsupervised learning approach provides a valuable alternative for optical flow estimation and shows good results on the evaluated datasets + It is unsupervised, which is another good keypoint to bear in mind + The ablation study on the loss terms impact and the comparison between supervised vs. unsupervised are very valuable

Weaknesses

- After reading the work, I don't think authors make clear the differences between spike and event cameras. There is a confusion all over the introduction and methodology and this makes me (and the reader) hesitate about the relevance of this work. - The work needs extensive English proofreading, some sentences are hard to understand. - The conclusions should not be a summary of the work - In the experimental section, authors compared on PHM and propose their own dataset but they are not comparing to the widely used MVSEC and DSEC datasets (that are the most common for event-based processing). This will improve not only completeness but also the comparison to other works in the state of the art. - From the paper and since authors propose an artificially generated dataset, it is not clear if 1) the spike camera is a real sensor and 2) in case it is, if they have access to real sequences recorded with it.

Questions

- Regarding the difference between event and spike cameras. Authors should make clear what kind of data they are handling. Event cameras produce asynchronous events and in fact, one of the advantages of these cameras is the lack of need for reconstructing the dynamics of the scene due to their high temporal resolution. This is great e.g. for optical flow estimation. However, it seems that spike cameras produce frames at very high resolution (something closer to a high-speed camera). This could also be interesting but it carries the burden of analyzing a lot of full frames, which cannot be ignored as a drawback. I really think that, since they are comparing to both conventional and event cameras, this should be clarified as soon as possible in the paper. In fact, this is mixed up in the paper, for example, at some point authors mention something about spike frames but then, they point out that pixels respond asynchronously (?) - When comparing to event camera processing, performance should be part of the comparison as well (or latency). - In the abstract, lns 12-14 do not make reference to the dataset (I think this should be added here as well) - In page 2, I believe the correct term is luminance (not illuminance) - In line 68, authors mention "shows visually impressive performance", what do you mean? Is that qualitative results? I guess authors mean that they are appealing? Are they only qualitative? - The sentence in Lns 107-108 seems incomplete - In Ln 182, are 40 and 100 ms? or frames? - In ln. 196 "different numbers of intervals" - Lsmooth is not defined - There are some recent works that use SNNs such as "Optical flow estimation from event-based cameras and spiking neural networks", from J. Cuadrado et al. that could help to improve the state of the art section. - Citations are weirdly formatted. Authors should make sure that they are using the conference format - In section 5.3 authors (unsupervised loss) authors mention Figure 4 when they mean Table 4. - What is the contribution of the LA module? According to Fig. 6 it seems it is not helping at all ... in such a case, why do authors keep it? is there another case in which the LA module does add a significant contribution? - At the end of section 5.4, authors discuss results on SSES dataset but they mention before that they are also doing some real-world sequences, where are the results on these sequences?

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

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

The work includes a limitations section that seems adequate

Reviewer 2gct4/10 · confidence 5/52023-06-29

Summary

This paper propose an unsupervised learning framework for spike-based optical flow estimation, which is mainly developed for spike input representation and spike loss function. In general, I think the method has merit, but the experiment is unconvincing in its lack of clarity.

Strengths

1. Propose a lightweight end-to-end spike data representation with the function of a dynamic time window. 2. Propose a two-stage unsupervised loss to model light intensity in regions with different motion speeds.

Weaknesses

1. In Fig. 5, there are significant quantization/discontinuities in the optical flow gt of the SSES dataset, concentrated in the ground area. To my knowledge, this is not present in other optical flow datasets. The reliability of the optical flow ground-truth is a concern, please provide an explanation. 2. In Fig. 4 and supplementary video, the results of these methods using only spike as input look bad. The major problem is that the results are concentrated only the textures and are not consistent for the same object., which cannot be called ``dense`` optical flow. To my knowledge, this phenomenon is not present in unsupervised two-frame based optical flow estimation methods. I suggest comparing the outputs of two-frame based methods and analyzing the spike-based failure cases. 3. The performance of using spike cameras for optical flow estimation is of concern. In L27, spike camera can record details of high-speed moving objects, it has enormous potential for estimating more accurate optical flow in high-speed scenes. But comparing the state-of-the-art optical flow estimation methods using image cameras, I believe this claim is not convincing. The visualization results of existing spike-based methods, including the authors', are significantly worse than existing image-based methods.

Questions

1. L158, Blindly fusing the output of all layers may impair the learning process. It is desirable to have experiments to support this belief. 4. L170 seems duplicate to L122. 5. I suggest to complete the experiments of raft backbone model.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

2 fair

Presentation

3 good

Contribution

3 good

Limitations

1. L132, advantage of neuromorphic vision is low latency. However, the authors do not present the runtime analysis of the proposed method. 2. See weaknesses above.

Reviewer eenq6/10 · confidence 3/52023-07-03

Summary

This paper works on the optical flow estimation problem, especially in the unsupervised setting, for high-frequency spike camera inputs. Specifically, a dynamic timing representation module and a spike-based unsupervised loss are proposed to improve performance. A new synthetic dataset, namely SSES, is created to test extreme scenes in autonomous driving. Experiment results show the proposed method has achieved state-of-the-art performance compared with previous methods.

Strengths

1. The paper is well-motivated by introducing the importance of optical flow estimation for high-frequency inputs collected from latest developed sensors like spike cameras. This topic has not been extensively studied so far. 2. The paper is overall clear and well-structured. 3. The proposed method is reasonable. Explanations on each module design are provided. 4. The results are compelling.

Weaknesses

1. There are still some confusions in the text or figures that need to be clarified. See additional comments below. 2. It will be better to polish the language of the paper before publication. Correct grammatical errors and use more formal language. Avoid informal expressions like "a bunch of" (Line 138), "won't" (Line 140), etc. See additional comments below. 3. Authors are encouraged to share their code and generated datasets for better reproducibility. Such plan has not been declared in the paper.

Questions

1. If I understand it correctly, the method uses the predicted optical flow to determine whether each pixel belongs to a low-speed or high-speed region and then uses different ways to approximate light intensities $\tilde{I}$. At the starting iterations of training, the predictions could be totally off, so it is likely that the method will pick a wrong way to approximate intensities. Did you try to tackle this issue? 2. The data used in Fig 4 do not have ground-truth labels, right? Only qualitative evaluation is possible, so this evaluation may be weak. 3. There is also a large part of supervised results, so stating "unsupervised" in the title could be misleading. 4. I am curious how a normal RGB-based optical flow network would perform as a comparison. For example, a naïve approach is to first convert spike inputs to RGB and then apply a state-of-the-art RGB-based optical flow network of similar size (parameters, memory cost). What would the results be like? This could be a baseline to strengthen the motivation of developing optical flow networks specific to spike camera inputs. Additional comments: 1. Grammatical errors: run-on sentences (Line 103-105, line 140-141, Line 163-165, Line 310-311). 2. Eq 1: The "mod" operator is traditionally only defined for integer division, but the value here is clearly continuous numbers, so it is better to define "mod" in the text in the continuous context. 3. Chapter 3.1: There are some confusions here. (1) I assume $A(\mathbf x, t)$ is continuous, but Line 119 says the output stream is binary, so there is missing gap in between that need explanations. Maybe adding a definition, something like $S_t^N(\mathbf x) = \mathbb I(\sup_{\tau\in[t-\Delta t, t+\Delta t]} A(\mathbf x, \tau) =\theta)$ will make it clear. (2) Line 113-114 say the camera "fires a spike" when the cumulative electrons hits the threshold, so I assume it is asynchronous for each pixel, but then Line 117-118 say the camera "fires a spike" at time T, which is synchronous and periodic. Maybe you could change the expression here to avoid confusion. 4. Line 130: "officially" -> "official". 5. Line 135: "dependence" -> "dependency". 6. Fig 2: Text is too small. It is better to make text size close to that of the main paper so that readers can read the figure clearly without needing to zoom in. 7. Line 147: It is better to add that your convolution is 1D on the time dimension to avoid confusion. 8. Line 182-183: Does your method guarantee $\omega_s + \omega_l = 1$? Eq 6: Does $\sum_k\omega_k=1$? If so, state so explicitly. 9. Eq 7 could be merged into the text. It does not have to an equation. 10. Line 175-184: What do you mean by "speed"? The frequency of camera readings, or fast motion? Please clarify. 11. Line 207: The notation $\omega_s, \omega_l, \omega_{k=1}, \omega_{k=2}$ is messy. Try not to use both textual abbreviations ("s", "l") and values ($k$) as subscripts for the same letter $\omega$. Use other letters like $\alpha_s, \alpha_l$, $\beta_k$ instead; or, you could do superscripts with parentheses $\omega^{(s)}, \omega^{(l)}, \omega_k$. 12. Line 209: "approximate" -> "approximated". 13. Line 212: define or cite "Charbonnier loss"; Line 213: define or cite "smoothness loss" specifically. There are many ways to define "smoothness". 14. Table 1 and 2: It is better to add citations into the table, so we know which data are generated from your experiments and which data are cited from other papers. The current table assumes that all data are credited to yourself. 15. Table 1 and 2: I recommend adding a vertical line between the "Top" and "Mean" columns, so we know "Top" is the name of the scene instead of the top error across all scenes. 16. Line 277: "through" -> "though". 17. Line 284: "Figure 4" -> "Table 4".

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

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

3 good

Limitations

Yes

Reviewer fVdD6/10 · confidence 3/52023-07-05

Summary

The paper introduces a method for optical flow estimaiton for spike cameras. Because of the high temporal resolution of the camera, a preprocessing step based on temporal dilated convolution and attention layers is used to automatically select the best temporal scale for a given sequence. The authors also introduce a unsupervised loss based on the reconstruction of intensity values form spike data combined with the photometric loss.

Strengths

The main strengths of the paper are the temporal dilated convolution preprocessing that allow for automatic temporal scale selection, as well as the unsupervised loss. The description of the method is clear and the experimental section complete, including ablation studies and comparision with previous works.

Weaknesses

The paper does not address the question of latency and computational time, which are key aspect of the spike camera. Also, since frames are available in the simulate datasets it would have been interesting to compare the accuracy of the proposed method on standard rgb frames, to quantify the benefits of using a high temporal resolution sensor for optical flow estimation. Finally, for people not familiar with the spike camera, it would be beneficial a more complete introduction to this type of sensor. There is no reference in the paper to any publications explaining what a spike camera is, it's HW implementation and carachteristics.

Questions

please refer to "weaknesess" section

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors could add a general discussion of advantages and limitations of the spike camera and of optical flow from spike cameras. For example any the challenge related to data rate, computational time and power consumption of these devices.

Reviewer PCU52023-08-10

Thanks for the responses

I really appreciate the responses. The authors have successfully addressed all of my concerns. While I am happy to raise the score, I have to stress that the paper focuses more on an engineering problem rather than a generic algorithmic challenge. I think this manuscript can be a very strong ICRA/IROS paper but may not provide much value in AI community. With that said, my conclusion is weak acceptance.

Reviewer PCU52023-08-11

Also, I think it would be very interesting to theoretically compare it to a state-of-the-art method, such as 'Transflow: Transformer as a Flow Learner [CVPR 2023]'. This may provide insight for the proposed method in relation to concurrent work.

Authorsrebuttal2023-08-11

Thank your for your advice

TransFlow is a related and interesting work. We will discuss TransFlow in related works. The idea of utilizing the long-range temporal association to recover more information ($\textit{e.g.}$, occlusion) in TransFlow provides valuable insights into addressing the occlusion issue within spike-based optical flow estimation. However, there are some differences between TransFlow and USFlow. The TransFlow focuses on improving the accuracy of correlations to estimate high-quality optical flow. While based on the characteristics of spike data, USFlow focuses on effectively and accurately representing spike data and designing a new unsupervised loss in spike-based optical flow estimation. From the perspective of data flow, there are differences in the strategies for constructing and utilizing correlations between TransFlow and USFlow. The TransFlow additionally uses the long-range temporal associations and utilizes Multi-head Self-Attention (MSA) and Multi-head Cross-Attention (MCA) to construct the correlations. Then it estimates the optical flow from the correlations. USFlow firstly represents spike data by the proposed Dynamic Timing Representation (DTR) module. Then it uses CNNs to construct correlations from these spike representations. Eventually, USFlow gradually refines the estimated optical flow under the guidance of correlations.

Reviewer Hpzy2023-08-12

Thanks for the rebuttal

I thank the authors for their rebuttal. Now, I have a clearer picture of the work. I think authors should carefully proofread the work for the final version. Also, and regarding the mentioned SNN-based works, I suggest some recent ones such as: - Optical flow estimation from event-based cameras and spiking neural networks - Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow

Reviewer 2gct2023-08-12

I have read the authors' responses as well as the comments of other reviewers. I agree that the methods are somewhat contributing, but I still maintain that the evaluation presented in this paper is limited, so I will lower my rating. I believe that all three of the weaknesses that concerned me have not been definitively addressed by the authors. For Q1, is the simulated optical flow reliable? The authors mentioned the limited quantization accuracy of CARLA as a problem, so does this problem still exist with the optical flow data stored in the simulated dataset? I think a conclusive answer is needed. For Q2, the authors start the paper by claiming that the spike camera has enormous potential for estimating more accurate optical flow in high-speed scenes (from line 26). But the results presented in their paper fall short of my expectations for the performance of state-of-the-art optical flow estimation methods. The authors do not answer why the optical flow estimated from spike data is not spatially dense. For Q3, the reason given by the authors for not comparing the two-frame method is the noise contained in the images, which does not convince me. From Figure 4, I think the image quality is acceptable and I believe that the existing two-frame methods can easily provide better optical flow estimation results. In addition, I think 90ms is still slow for low latency neuromorphic vision, but of course this is not a problem that should be addressed by every frontier study.

Authorsrebuttal2023-08-12

Thank you for your response

We will give clearer clarifications to your concerns below. For the concern of Q1, we checked the unique values of the gt, sorted them, and looked at the diff between each unique value. The diff is exactly 1/1024 which points to a 10-bit quantization. Thence, the dataset is reliable, and its accuracy is 1/1024. We will release this dataset for others to check. For the concern of Q2 and Q3, we would change the claim to $\textit{it has enormous potential for estimating optical flow in high-speed scenes}$. We admit some existing two-frame methods can provide better optical flow estimation results. However, we want to clarify that when the motion to be estimated occurs during an extremely short time, the conventional RGB camera cannot provide the corresponding RBG frames due to its low frequency. In other words, If we cannot obtain data in some high-speed scenarios, it is hard to obtain optical flow. Hence, the spike camera and conventional RGB cameras have different application scenarios. Spike camera is only a complement for the various vision tasks. Since it is not a replacement for RBG camera, it cannot surpass the frame-based work in all scenarios. In addition, as a counterpart, event camera also cannot surpass the RBG camera in optical flow estimation if cameras both have data for a certain scenario. However, it doesn’t mean there is no meaning to investigating this direction and it is indeed a thriving field at the current stage. The unsupervised finetune dataset is small, which leads to the phenomenon that the estimated optical flow is not spatially dense. Note that the results are spatially dense in Figure 3. As mentioned by other reviewer, unsupervised spike-based optical flow estimation is a new venue. However, the unsupervised frame-based methods have been studied for a long age, and they have many better training datasets which can boost the generalization. We admit at the current stage, our method has limitations, but our results look not bad and are supported by other reviewers. In addition, only when this new venue is encouraged, more robust datasets and more works can emerge to promote the development of this field.

Reviewer PCU52023-08-12

Why lower your rating?

Reviewer 2gct, Thank you for your diligent review efforts. I am currently engaged in a thorough examination of your comments and find myself engaged in a process of comprehending the underlying rationale behind your suggestion to "lower my rating." Should the authors have encountered challenges in adequately addressing your inquiries, it is reasonable to consider our initial evaluation as the benchmark. Conversely, any diminution of the rating is justifiable only in cases where the authors have, inappropriately, provided responses that are contrary to the essence of your inquiries. Upon careful scrutiny, I am inclined to assert that the authors have, to the best of my observation, diligently addressed the issues raised in your comments, offering responses that are in alignment with the nature of your inquiries. Consequently, it is my contention that the grounds for reducing the rating appear to lack a sufficiently robust foundation. It is essential to underscore the significance of fostering a constructive discourse within the academic community to facilitate the cultivation of insightful new knowledge. As many of us are actively engaged in various capacities, encompassing roles as authors, reviewers, and ACs, we bear a collective responsibility to uphold the principles of impartial and judicious assessment that are integral to the progress of our academic community. This communication serves as a respectful reminder of our shared commitment to the responsibilities inherent in the role of a reviewer. Its intent is in no manner geared towards meddling with the established rating, but rather serves as a courteous prompt to reiterate our shared role in promoting a culture of conscientious and equitable evaluation. Thanks for reading my message.

Reviewer fVdD2023-08-16

Thank you for your answers. I still believe that the points raised in the initial review are relevant and their discussion should be added to the article.

Authorsrebuttal2023-08-16

Thank you for your response

Thank you for your suggestions. The points you raised in the initial review are important, and we will add these discussions to the paper as you suggested.

Authorsrebuttal2023-08-21

Thanks to reviewers

We express our sincere gratitude to reviewers for their constructive suggestions. We will certainly add the corresponding discussions to this paper as reviewers suggested. Thank you for your approval of this work.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC