Abstract
DOTA (Dataset for Object deTection in Aerial images) is a comprehensive dataset designed to advance object detection in Earth Observation. It comprises 188,282 instances across 15 diverse categories, annotated with oriented bounding boxes. This approach addresses unique aerial imagery challenges, including arbitrary object orientation, vast scale variations, and high spatial density. By providing a large-scale, expertly annotated benchmark, DOTA bridges the gap between general computer vision and remote sensing, offering a rigorous platform for developing robust aerial object detectors.
Introduction
Most conventional object detection datasets, such as MSCOCO or ImageNet, are built on "natural scenes" where objects typically follow a vertical orientation due to gravity. Consequently, standard detectors rely on axis-aligned (horizontal) bounding boxes. However, aerial imagery captured from satellites or drones presents a different paradigm. From a bird's-eye view, objects like ships, planes, and vehicles can appear at any orientation.
Aerial detection faces three primary hurdles:
- Scale Variation: Objects range from tiny cars to massive harbors within the same dataset.
- Arbitrary Orientation: Objects are rarely aligned with image axes.
- High Density: In areas like parking lots or docks, horizontal boxes overlap significantly, leading to false suppressions by standard algorithms. DOTA addresses these by introducing oriented bounding boxes (OBB).
Methods
The DOTA framework utilizes an Arbitrary Quadrilateral Annotation scheme. Instead of the standard 4-parameter box $(x, y, w, h)$, DOTA uses 8 parameters representing the $(x, y)$ coordinates of four vertices. This allows for tight fitting around tilted objects. Annotations were performed by experts to ensure high quality.
To establish baselines, the authors adapted several state-of-the-art detectors (Faster R-CNN, R-FCN, YOLOv2, SSD). A key modification was the "Faster R-CNN Oriented" (FR-O), which adjusts the Region Proposal Network (RPN) and regression heads to handle 8-degree-of-freedom quadrilaterals. Due to the massive size of aerial images (up to 4000x4000 pixels), a specialized pipeline crops them into 1024x1024 overlapping patches for processing.
The DOTA Dataset
DOTA is significantly more diverse and challenging than previous benchmarks like NWPU VHR-10. It contains 2,806 images and 188,282 instances across 15 categories.
The categories include: Plane, Ship, Storage Tank, Baseball Diamond, Tennis Court, Swimming Pool, Ground Track Field, Harbor, Bridge, Large Vehicle, Small Vehicle, Helicopter, Roundabout, Soccer Ball Field, and Basketball Court.
Key Insights and Results
A major finding is the failure of standard Non-Maximum Suppression (NMS) in crowded scenes. Horizontal boxes for adjacent ships often overlap almost entirely, causing NMS to delete valid detections. Oriented boxes solve this by maintaining distinct boundaries. Furthermore, the authors noted that spatial resolution (meters per pixel) is crucial in aerial imagery for distinguishing objects of different physical sizes that look similar.
Key results include:
- Benchmark Difficulty: Baseline models performed significantly worse on DOTA than on natural scene datasets, confirming its complexity.
- OBB Superiority: The oriented Faster R-CNN achieved an mAP of 54.13%, outperforming horizontal models, especially for high-aspect-ratio objects like ships.
- Generalization: Models trained on DOTA generalized better to other datasets than vice versa, indicating DOTA's role as a comprehensive "superset" for the aerial domain.
- Small Object Focus: Over 57% of instances are "small" (10-50 pixels), making DOTA an ideal benchmark for small-object detection research.
Conclusion
DOTA sets a new standard for aerial object detection. While it faces challenges like class imbalance and the ambiguity of "stuff" categories (e.g., defining harbor boundaries), it provides the necessary scale and complexity for future research. Future work will likely focus on optimizing the computational efficiency of oriented detectors and addressing the long-tail distribution of object classes.