What if CVPR is a Graphics Conference?
At the CVPR 2020 workshop on AI for Content Creation, I gave an informal talk about the increasing numbers of computer graphics (image synthesis) papers appearing at computer vision conferences:
Here’s a summary of the main points:
- Some of the most exciting work in image synthesis is appearing at computer vision conferences
- This is suprising, since SIGGRAPH has always been THE place to publish the best computer graphics work
- See also this post and the comments on this Facebook thread for discussion about why.
- The way research is conducted and evaluated is a product of the interests of the members of the community
- My stereotype of typical computer vision researchers is that they aim to solve “the AI problem”, and are focused on automation and quantiative evaluation.
- My stereotype of typical computer graphics researchers is that they aim to create tools that help create art, creative expression and visual communication. They value tools that provide user control and provide high visual quality.
- As a result, computer vision has some blind spots for graphics topics
- There is an over-emphasis on quantitative evaluation, even when that evaluation isn’t meaningful
- Visual qualities of the results are not sufficiently valued, or emphasized as a research goal
- Too much research focuses on variants of deep learning models without attempting to think about the actual visual phenomenon being modeled.
- It’s exciting to see all this great progress, taking ideas from both fields
- Many of the best image synthesis papers in vision conferences come (at least, in part) from authors with graphics backgrounds