Adaptive parameter tuning for morphological segmentation of building facade images

We present an adaptive method to segment Haussmannian facades from street level images. Our approach assumes that images are rectified, cropped and their elements are aligned in a pseudo-regular structure. It is based on the accumulation of directional color gradients, combined with morphological filters in order to deal with textured facades. We propose an automatic parametrization of three filters included in the process: opening filter of size n_op, alternate sequential filter (ASF) of size n, and H-minima filter with contrast threshold h. This automatic selection offers robustness to noise, image resolution changes, shadows and textures. Quantitative and qualitative results are reported on a public annotated database, validating the good performances of our approach.