UXOTi_NPA_11_class dataset

Evaluation dataset <strong>UXOTi_NPA</strong>–an abbreviation that stands for UneXploded Ordnance Thermal images provided by Norwegian People's Aid–for this research was constructed as follows. The raw data was provided by the humanitarian organization Norwegian People’s Aid which collected videos about unexploded ordnances during a project in Bosnia and Herzegovina in 2019. Thermal imaging sensor was mounted on the UAV that surveyed the area at a height of 3 meters. A very small Ground Sampling Distance (GSD) was achieved with such a configuration. These videos were acquired by the thermal imager Zenmuse XT with a dimension of 720x480 pixels at 30 Hz frame rate. Pixel pitch was 17 μm, sensitivity (NEdT) &lt;50mK at focal length f/1.0, and a field of view was 69°x56°. Videos were stored in mp4 format. Image sequences were grabbed from the full videos by using VirtualDub software. These clips were about 25 seconds long and show the UAV's flight over the targets. We annotated videos semi-automatically by using Computer Vision Annotation Tool (CVAT). Positions of objects’ annotations were manually corrected after every 5 to 6 frames. The evaluation dataset consists of 808 thermal images of dimensions of 720x480 pixels extracted from the above-described video clips, whenever at least one target (UXO) was present in the image. The number of objects per image varied between one to three.