SODwS-V1
Description
Dataset for Small Object Detection with Shadow (SODwS)
Files
Steps to reproduce
A Sony HDR-TD30V digital video camera was used for image acquisition across two scenes. In the first scene, the camera was positioned on the 4th floor of a faculty building. For the second scene, it was set up on the highest bench of the stadium grandstand. The objective was to capture aerial images showcasing shadows and small objects influenced by these shadows. The image acquisition setups were classified into four categories: NSNO (No Shadow, No Object), NSWO (No Shadow, With Object), WSNO (With Shadow, No Object), and WSWO (With Shadow, With Object). The recorded video was manually reviewed to select frames fitting any of these categories. For each selected image, a binary shadow mask was manually created. Small objects annotated included bags, balls, boxes, sport cones, hats, bike helmets, bike pumps, and sneakers. Annotations were provided in YOLOv9 format, with each object’s region defined by the coordinates of the upper-left and lower-right corners of a rectangular bounding box. Image augmentation techniques were applied to images containing small objects, resulting in a total of 385 aerial view images. Both image annotation and augmentation were performed using Roboflow.