Weakly Supervised Semantic Segmentation Learning on UAV Video Sequences

The domain of scene understanding from Unmanned Aerial Vehicles (UAVs) is of high interest for researchers in the computer vision domain, since it can be used for object detection and tracking in scenarios like deforestation monitoring, traffic surveillance, or for civil engineering tasks. However, the topic of dense video segmentation from drones has been insufficiently explored due to the lack of annotated ground truth data. We propose a solution based on a framework composed of a deep neural network for semantic segmentation and an optical flow generator, linked together by a spatio-temporal GRU component to efficiently solve the problem of weakly supervised semantic segmentation of video sequences recorded from UAVs. The novelty of our work comes from the employment of depthwise separable convolutions for the GRU component, which decrease the computation time and increase the segmentation accuracy. We test our methodology on the synthetic dataset Mid-Air, for low-altitude drone flight, and report results that prove the usefulness of the proposed system.