Event Location: NSH 3305
Bio: Olga Russakovsky (http://ai.stanford.edu/~olga) is a PhD student at Stanford University advised by Professor Fei-Fei Li. Her main research interests are in large-scale object detection and recognition. For the past two years she has been the lead organizer of the international ImageNet Large Scale Visual Recognition Challenge which has been featured in the New York Times, MIT Technology Review, and other international media venues. She has organized several workshops at top-tier computer vision conferences: the ImageNet challenge workshop at ICCV’13 and ECCV’14, the upcoming workshop on Large-Scale Visual Recognition and Retrieval at CVPR’15, and the new Women in Computer Vision workshop at CVPR’15. During her PhD she collaborated closely with NEC Laboratories America and with Yahoo! Research Labs. She was awarded the NSF Graduate Fellowship and the CRA undergraduate research award.

Abstract: There are four key components when tackling large-scale visual recognition: (1) the data needs to have sufficient variety to represent the target world, (2) the algorithms need to be powerful enough to learn from this data, (3) the data needs to be annotated with enough information for algorithms to learn from, and (4) the algorithms need to be fast enough to process this large data. I’ll talk about my PhD work (including an upcoming CVPR’15 paper) on each of these four components in the context of large-scale object detection.

The main part of the talk will focus on the difficulties of collecting and annotating diverse data when designing the object detection task of the ImageNet Large Scale Visual Recognition Challenge (http://image-net.org/challenges/LSVRC/). ILSVRC is a benchmark in object classification and detection on hundreds of object categories and millions of images. The challenge has been run annually since 2010, attracting participation from more than fifty institutions. In 2014 the challenge had a record number of submissions (123 entries from 36 teams) and appeared in international media including the New York Times, MIT Technology Review and CBC/Radio-Canada. In this talk I’ll describe some of the challenges of scaling up the object detection task of ILSVRC by more than an order of magnitude compared to previous datasets (e.g., the PASCAL VOC).

I will conclude by discussing my current work and future plans. As computers are becoming exceedingly good at recognizing many object categories, I argue that it’s time to step back and consider the long tail: what pieces are still missing and preventing us from recognizing /every/ object and understanding /every/ pixel in an image? I believe that we’ll need to closely examine the performance of our algorithms on individual classes (instead of focusing on average accuracy across hundreds of categories), revisit the idea of object description rather than categorization, and consider the advantages of close human-machine collaboration.