Performance Analysis of Semantic Segmentation Algorithms for Finely Annotated New UAV Aerial Video Dataset (ManipalUAVid)

Semantic segmentation of videos helps in scene understanding, thereby assisting in other automated video processing techniques like anomaly detection, object detection, event detection, etc. However, there has been limited study on semantic segmentation of videos acquired using Unmanned Aerial Vehicles (UAV), primarily due to the absence of standard dataset. In this paper, a new UAV aerial video dataset (ManipalUAVid) for semantic segmentation is presented. The videos have been acquired in a closed university campus, and fine annotation is provided for four background classes viz. constructions, greeneries, roads, and waterbodies. Also, the performance of four semantic segmentation approaches: Conditional Random Field (CRF), U-Net, Fully Convolutional Network (FCN) and DeepLabV3+ are analysed on ManipalUAVid dataset. It is seen that these algorithms perform competitively on UAV aerial video dataset and achieves an mIoU of 0.86, 0.86, 0.86 and 0.83 respectively.