TinyPedSeg: A Tiny Pedestrian Segmentation Benchmark for Top-Down Drone Images

Abstract

The usage of Unmanned Aerial Vehicles (UAVs) has significantly increased in various fields such as surveillance, agriculture, transportation, and military operations. However, the integration of UAVs in these applications requires the ability to navigate autonomously and detect/segment objects in real-time, which can be achieved through the use of neural networks. Despite object detection for RGB images/videos obtained from UAVs are widely studied before, limited effort has been made for segmentation from top-down aerial images. Considering the case in which the UAV is extremely high from the ground, the task can be formed as tiny object segmentation. Thus, inspired from the TinyPerson dataset which focuses on person detection from UAVs, we present TinyPedSeg, which contains 2563 pedestrians in 320 images. Specialized only in pedestrian segmentation, our dataset presents more informativeness than other UAV segmentation datasets. The dataset and the baseline codes are available at https://github.com/ituvisionlab/tinypedseg