Open ECCV Paper Review Comment

[Kin-Zhang]

To let everything open and public, I will public all review comments I got in this thread, feel free to check and have your own opinion.

And you can check all comments and our rebuttal file by click here to view the PDF

[Kin-Zhang]

Reviewer 1:

14. Final rating based on ALL the reviews, rebuttal, and discussion (post-rebuttal).
    Strong Accept

15. Final justification (post-rebuttal).
    I maintain my rating of a strong accept. Importantly I disagree with R2 and R3's criticism about pre-computing dynamic points being a weakness of this method. Classifying static vs dynamic points is an important part of the scene flow problem and designing a better method for self-supervising this portion of the problem is very much in the spirit of a better self-supervised scene flow method.

[Kin-Zhang]

Reviewer 2:

14. Final rating based on ALL the reviews, rebuttal, and discussion (post-rebuttal).
    Borderline Accept

15. Final justification (post-rebuttal).
    Thank you for the rebuttal.
    I have read the reply one by one. My responses are as follows:
    1). Performance comparison: The explanation clarifies my concern on performance. Thank you.
    2). Loss: I see the value of eq 11 in handling the upper bound f_tilde. Thank you for explaining.
    3). Loss: unfortunately, my opinion remains: using offline pre-calcuated dynamic/static points to aid scene flow introduces extra noise in learning, which may not be an optimal solution to scene flow tasks whose goal is to predicting per-point motion or distinguish dynamic/static, although this strategy indeed improves the overall result. The authors seem to have a strong opinion on this comment.

One additional remark on the rebuttal, I find it hard absorb the condensed rebuttal. I would appreciate point-wise reply in paragraphs. Clarity matters.

Overall, the paper has nice observations and proposes additional loss functions to handle the observed issues. The experimental results are good as well. However, as I have explained in the final justification, I do not think "having precomputed static/dynamic points" is a good strategy for scene flow tasks whose end goal is to separate the dynamic from the static by per-point motion. My final rating is borderline.

[Kin-Zhang]

The additional remark I'd like to say: 1. the initial comment this reviewer had at least 15-18 points/question. 2. Then it's impossible to do point-wise reply inside 1-page rebuttal limitation. 😇

[Kin-Zhang]

Reviewer 3:

14. Final rating based on ALL the reviews, rebuttal, and discussion (post-rebuttal).
    Weak Accept

15. Final justification (post-rebuttal).
    Thank you for the rebuttal. I agree with Reviewer 2 about the limitations of using precomputed static/dynamic points as a prerequisite of the method. I think this shortcoming should be discussed in detail in a discussion/limitations section. However, I still find the study valuable with strong empirical results. The community will benefit from the results reported in this paper.

My final rating is weak accept.