Object recognition in small images
R. Fergus, A. Torralba
The human visual system is remarkably tolerant to degradations
in
image resolution: in a scene recognition task, the performance of
subjects is similar whether 32x32 color images or multi-mega pixel
images are used. Accordingly, we consider the task of
scene recognition in computer vision using very small (32x32 pixel)
images. Small images force us to employ contextual reasoning and to explore
new representations due to the lack of textural cues on which many existing
algorithms are based. However, the small size of each image carries two
important benefits: (i) it permits standard machine learning tools to be
easily applied and (ii) huge image databases may be easily collected.
Inspired by the data-driven nature of Google, we deploy simple data
mining methods to a large collection of images (50 million). We will
show very preliminary results.
presentation