Where machine vision needs help from machine learning

I'll describe where computer vision needs advances from computer science and machine learning. This talk will cover where computer vision works well: finding cars and faces, operating in controlled environments, and where it doesn't work well: in the uncontrolled settings of daily life. Several aspects of the problem make it particularly appropriate for machine learning research: we have large datasets of high-dimensional data, so efficient processing is crucial for success. The data are noisy, and we search and analyze images over Internet scales. I'll list a number of computer vision problems, describe their structure, and tell where we need help. This talk was partially crowd-sourced: at recent computer vision conferences, I've asked my colleagues where they felt we needed help from computer science and machine learning, and I'll report on what they said.