People detection in fish-eye top-views

Is the detection of people in top views any easier than from the much researched canonical fronto-parallel views (e.g. Caltech and INRIA pedestrian datasets)? We show that in both cases people appearance variability and false positives in the background limit performance. Additionally, we demonstrate that the use of fish-eye lenses further complicates the top-view people detection, since the person viewpoint ranges from nearly-frontal, at the periphery of the image, to perfect top-views, in the image center, where only the head and shoulder top profiles are visible. We contribute a new top-view fish-eye benchmark, we experiment with a state-of-the-art person detector (ACF) and evaluate approaches which balance less variability of appearance (grid of classifiers) with the available amount of data for training. Our results indicate the importance of data abundance over the model complexity and additionally stress the importance of an exact geometric understanding of the problem, which we also contribute here.