Abstract:

This paper describes the use of two different deep-learning approaches for object detection to recognize a toy soldier. We use recordings of toy soldiers in different poses under different scenarios to simulate appearance of persons on footage taken by drones. Recordings from a bird's eye view are today widely used in the search for missing persons in non-urban areas, border control, animal movement control, and the like.
We have compared the single-shot multi-box detector (SSD) with the MobileNet or Inception V2 as a backbone, SSDLite with MobileNet and Faster R-CNN combined with Inception V2 and ResNet50. The results show that Faster R-CNN detects small object such as toy soldiers more successfully than SSD, and the training time of Faster R-CNN is much shorter than that of SSD.