Estimating depth from a single image represents an attractive alternative to more traditional approaches leveraging multiple cameras. In this field, deep learning yielded outstanding results at the cost of needing large amounts of data labeled with precise depth measurements for training. An issue softened by self-supervised approaches leveraging monocular sequences or stereo pairs in place of expensive ground truth depth annotations. This paper enables to further improve monocular depth estimation by integrating into existing self-supervised networks a geometrical prior. Specifically, we propose a sparsity-invariant autoencoder able to process the output of conventional visual odometry algorithms working in synergy with depth-from-mono networks. Experimental results on the KITTI dataset show that by exploiting the geometrical prior, our proposal: i) outperforms existing approaches in the literature and ii) couples well with both compact and complex depth-from-mono architectures, allowing for its deployment on high-end GPUs as well as on embedded devices (e.g., NVIDIA Jetson TX2).

Abstract:The main goal of this paper is to detect roads from aerial imagery recorded by drones. To achieve this, we propose a modification of SegNet, a deep fully convolutional neural network for image segmentation. In order to train this neural network, we have put together a database containing videos of roads from the point of view of a small commercial drone. Additionally, we have developed an image annotation tool based on the watershed technique, in order to perform a semi-automatic labeling of the videos in this database. The experimental results using our modified version of SegNet show a big improvement on the performance of the neural network when using aerial imagery, obtaining over 90% accuracy.