[link]
Summary by Hugo Larochelle 5 years ago
Loading...
Thanks for the very useful summary. I have a question regarding the fourth paragraph (starting with the text - "As for object features, ...") of your note. Are the interpretations for 'object features' and 'contextual features' gone interchanged? I mean - will it not be that 'object features' are those features whose values tend to change a lot when computed on a complete original image versus when computed on an image whose regions within object bounding boxes are blackened out? In section 4.2.1 this scenario defines the 'object score' $s^o_l$, and is it not natural that object features are those for which object scores are high?

I think you're right! I just fixed the paragraph, thanks!

Thanks Hugo. I think there's a very small typo in the penultimate line of the last paragraph (just before "My two cents"). You wrote - the object score is always higher for the top causal features than for the top anticausal features. However, the paper finds that object scores are higher for top anticausal features than for top causal features (Fig 5 and sec 4.3). In the end, thanks again for the good summary. It helped me a lot understanding the paper on a topic on which I did not have prior experience.

Oh man, yeah, there too that's a typo. Thanks for pointing that one out too!

Your comment:


ShortScience.org allows researchers to publish paper summaries that are voted on and ranked!
About

Sponsored by: