[2022 spring] NeurIPS 2021 The Emergence of Objectness: Learning Zero-shot Segmentation from Videos (20226056)

[emjay73]

• Reference section is empty
• In the Method section, I somehow understand the proposed method, but I still do not understand what will be the final output of the proposed network. Will it be all 5 segmentation channels?
• I wonder what happens to the number of segments when the video contains many moving objects heading towards different directions.
• How is it possible to detect salient objects in static images using the proposed method?

[ChoiDM]

Thank you for sharing your insightful review.

There are several points that I noticed:

• References are omitted
• Replacing the equations to the center position would be better
• Increasing the fontsize of subsection title (e.g., Related work, Idea) would be better
• It would be better to make the font of titles in 3. Method section (e.g., Appearance pathway for feature-based segmentation) bold and remove the repeated numbering (all subsection have the first 1. numbering)

[Anderlaxe]

Very interesting review. I notices some typos

• Subsections of section 3 all have a head number of 1 instead of something like 3.1, 3.2 etc
• At the very end of section 3, the variables X_i and X_j are not written in an equation format which can confuse the reader.
• In the description of the experimental setup (section 4) the values of the learning rate and weight decay are missing.

There is no reference in the paper. It would be interesting to mention the articles used in the description of the related work.

The issues and main interests of this article are clearly defined. The content is concise and points to the essential parts of the paper.