Visual attributes for enhanced human-machine communication

In computer vision systems today, humans typically communicate with the machine via limited interactions e.g. providing coarse image labels. This seems rather wasteful because it is precisely the human abilities that we aim to replicate in automatic image understanding. Moreover, humans are often meant to interact with vision systems as users (e.g. image search) or as supervisors training the system - be it for niche applications or for generic visual concepts such as everyday objects and scenes. On the flip side, machines today also rarely communicate with humans. Vision models are often complex and non-transparent. They simply fail without explaining why which is frustrating for users and perplexing for researchers. Here we describe some of our recent efforts towards using attributes to enhance the mode of communication between humans and machines to improve visual recognition.