A Large-Scale Dataset for Small and Tiny Object Detection from Aerial Imagery

Aerial object detection has gained significant attention due to its broad applications in areas such as security, surveillance, and agriculture. However, existing datasets present several challenges, such as annotation errors and limited scale, which hinder the accuracy and speed of current detection models. To address this, the NITJ-AIR dataset - an open-source, large-scale dataset has been proposed that aims to advance research in small and tiny object detection. Baseline results are provided for several state-of-the-art detectors, including Faster R-CNN, SSD, YOLO (v3, v4, and v5s), and CornerNet. Further, the performance of these detectors on NITJ-AIR is compared against the VisDrone and AU-AIR datasets. The comparative analysis demonstrates enhanced feature extraction capabilities, yielding a substantial minimum improvement of 8.37% in average precision and 7.65% in mean average precision across models. The developed dataset is available at https://github.com/himanshugiriraj/NITJ-AIR.