Small Ship Detection on Optical Satellite Imagery with YOLO and YOLT

Abstract

Actually, the use of deep learning in object detection gives good results, but this performance decreases when there are small objects in the image. In this work, is presented a comparison between the last version of You Only Look Once (YOLO) and You Only Look Twice (YOLT) on the problem of detecting small objects (ships) on optical satellite imagery. Two datasets were used: High-Resolution Ship Collection (HRSC) and Mini Ship Data Set (MSDS), the last one was built by us. The mean object’s width for HRSC and MSDS are 150 and 50 pixels, respectively. The results showed that YOLT is good only for small objects with 76,06% of Average Precision (AP), meanwhile, YOLO reached 69,80% in the MSDS dataset. Moreover, in the case of the HRSC dataset where have objects of different sizes, YOLT obtained a 40% of AP against 75% of YOLO.