CVPR Daily - Wednesday

Holger thinks that currently computer vision is a bit incomplete. He says we only really look at objects and have to develop much better models to take context into account. He thinks this work is a good new benchmark to evaluate new methods for these tasks. The most difficult part, he says, has been scaling the work up . It was easy when they did it for 10,000 images in a university with their own students annotating images. When they went up to 164,000 images, they had a company - Mighty AI - do the annotations for them, and once it’s crowdsourced, it gets much more complex. You need to do a lot of verification to ensure that the annotations are good. Finally, Holger wants to point out that they did not invent the stuff topic. It goes back decades. You can find it in psychology and linguistics and lots of other places. They are just one piece of the puzzle and many others are working on the same mission. “Currently, computer vision is a bit incomplete!” Wednesday 7 Holger Caesar