Title: On the road to understanding human and machine performance in object recognition

Abstract: According to current benchmarks machine performance on common object detection tasks approaches or even surpasses that of humans. Yet everyone's experience running object detectors indicates this is far from true. This has had an unfortunate effect on the community where because of saturation on such tasks we see little improvement recently and significant media hype. We'll talk about why this is and how to really know if machines are as good as humans. I'll present some early results. We're also looking for collaborations to make this as useful and scientifically meaningful as possible.