search

awesome-davian / awesome-reviews-kaist

Computer vision paper reviews written by KAIST AI students
Creative Commons Attribution Share Alike 4.0 International
43 stars 117 forks source link

[2022 Spring] BMVC 2021 M-CAM: Visual Explanation of Challenging Conditioned Dataset with Bias-reducing Memory (20214805) #475

Open MonumentalCloud opened 2 years ago

MonumentalCloud commented 2 years ago

Thank you for your work! It was a well detailed, however unfinished work. I was getting ready to be interested and the review abruptly ended.

  1. Some formatting issues need to be resolved, especially in the first section
  2. It would be nice to talk about general application in the real world as well
  3. Your Experimentation and Conclusion seems incomplete, it might be a pull request merge issue, I strongly encourage you to check it out.
  4. Also your method needs to be expanded, which right now only talks about Bias-reducing memory. How is that bias-reducing memory fit into the rest of the architecture?

Thank you for work again!

Junhyeon-Park commented 2 years ago

Thank you for review. Introduction and related work is good, but method and experiment part seems not finished. And formatting issuses in introduction make it difficult to read. I recommend that supplement the overall content including details.

YeodongYoun95 commented 2 years ago

Thank you for reviewing an interesting paper in class activation map. It was interesting that we can use such map to identify crucial regions for performing classification task, especially in the healthcare domain. The motivation for this paper was acknowledgeable as well, as in real datasets imbalanced dataset and multi-object multi-class problems often occur which hinders overall performance. In such situations, data augmentation, segmenting problems into smaller ones have been used in previous papers, but this paper proposed a novel bias-reducing memory module.

  1. The formatting and editing of the review overall feel unfinished and need more additional material and clipping. The experiment section is missing as well.
  2. It would be a lot helpful if more figures or equations are added such that the readers can understand the concept more directly. Why did CAM compute the weighted sum of feature values only at the last conv layer? Is there any mathematical intuition to it? What are the high-order derivatives of Grad-Cam they used to generalize Cam? It's really hard to understand and vague only by reading the text.

Thank you!