Entrance Detection from Street-View Images

We present a system for detecting building entrances in outdoor scenes, an important problem for urban scene understanding. While entrance detection in indoor scenes has received a lot of attention, tackling the problem in outdoor scenes is considerably more complicated and remains largely unexplored. The wide variety of door appearances and geometries, background clutter, occlusions, specularity, and other difficult lighting conditions together impose many challenges. In this paper, we propose a three stage system that starts with a high-recall entrance candidate extractor. The next stage classifies candidates based on local image features. The final stage fuses results from multiple views by using MCMC to solve a Bayesian inference problem, and to select the best set of entrances that explain the image of a facade. We achieve a precision of 70% at a recall of 70% on a challenging dataset of urban scene images. We will release this benchmark dataset to the public to facilitate future research on this topic.