Foveal vision for instance segmentation of road images

Instance segmentation is an important task for the interpretation of images in the area of autonomous or assisted driving applications. Not only indicating the semantic class for each pixel of an image, but also separating different instances of the same class, even if neighboring in the image, it can replace a multi-class object detector. In addition, it offers a better localization of objects in the image by replacing the object detector bounding box with a fine-grained object shape. The recently presented Cityscapes dataset promoted this topic by offering a large set of data labeled at pixel level. Building on the previous work of (Uhrig et al., 2016), this work proposes two improvements compared to this baseline strategy leading to significant performance improvements. First, a better distance measure for angular differences, which is unaffected by the -p/p discontinuity, is proposed. This leads to improved object center localization. Second, the imagery from vehicle perspective includes a fixed vanishing point. A foveal concept counteracts the fact that objects get smaller in the image towards this point. This strategy especially improves the results for small objects in large distances from the vehicle.