Object Detection in Aerial Images: A Large-Scale Benchmark and Challenges

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Object detection in natural images has advanced significantly, but aerial images present unique challenges due to scale and orientation variations.
• The lack of large-scale benchmarks hinders progress in object detection in aerial images (ODAI).

Purpose of the Study:

• To introduce a large-scale dataset, DOTA, for object detection in aerial images.
• To establish comprehensive baselines and facilitate reproducible research in ODAI.

Main Methods:

• The DOTA dataset comprises 11,268 aerial images with 1,793,658 object instances across 18 categories, using oriented-bounding-box annotations.
• Evaluated 10 state-of-the-art algorithms with over 70 configurations for speed and accuracy.
• Developed a code library and evaluation website for ODAI.

Main Results:

• The DOTA dataset provides a robust foundation for ODAI research.
• Extensive baselines offer performance insights into various algorithms.
• Previous challenges on DOTA attracted over 1300 international teams.

Conclusions:

• The DOTA dataset, baselines, and associated resources aim to accelerate the development of robust ODAI algorithms.
• Facilitates reproducible research and benchmarking in the field of object detection in aerial images.