In the main part of the talk, I will discuss how computers can be taught to count objects in images, such as cells in microscopy images or pedestrians in surveillance videos. This task is approached within the supervised machine learning framework, where a computer learns from limited amount of human-annotated images. I will present a learning procedure that uses a quality-of-fit (loss) function designed specifically for the counting task. In the experiments, this procedure achieved considerably higher counting accuracy than procedures using standard losses. I will then briefly discuss the ongoing project that uses the learning-to-count framework to estimate cell colony counts in low-resolution wide field-of-view microscopy. The talk will conclude with the discussion of potential future research directions in biomedical image analysis and large-scale visual recognition.
Victor Lempitsky currently holds a research position in computer vision at Yandex, Moscow. Prior to that, he was a postdoc researcher with the Visual Geometry Group at the University of Oxford and with the Computer Vision Group at Microsoft Research Cambridge. Victor holds a PhD degree (“kandidat nauk”) in applied mathematics (2007) from Moscow State University. His research interests are in visual recognition and biomedical image analysis. He has co-authored more than 20 papers in ICCV, CVPR, ECCV, NIPS, MICCAI conferences as well as TPAMI, IJCV, and JMIV journals. He is an area chair for the ECCV’2012 conference and has also received a best paper award at the international symposium on Functional Imaging and Modelling of the Heart (FIMH'2009).
Seminar host: Regina Barzilay