In the past decade, significant progress in object detection has been made in natural images, but authors of the DOTA v2.0: Dataset of Object deTection in Aerial images note that this progress hasn’t extended to aerial images. The main reason for this discrepancy is the substantial variations in object scale and orientation caused by the bird’s-eye view of aerial images. One major obstacle to the development of object detection in aerial images (ODAI) is the lack of large-scale benchmark datasets. The DOTA dataset contains 1,793,658 object instances spanning 18 different categories, all annotated with oriented bounding box annotations (OBB). These annotations were collected from a total of 11,268 aerial images. Using this extensive and meticulously annotated dataset, the authors establish baselines covering ten state-of-the-art algorithms, each with over 70 different configurations. These configurations are evaluated for both speed and accuracy performance.

Regarding the construction of DOTA, the authors emphasize the importance of collecting images from various sensors and platforms to address dataset biases. They describe the acquisition of such as Google Earth, the Gaofen-2 Satellite, Jilin-1 Satellite, and CycloMedia airborne images. These images vary in resolution and sensor type, reflecting real-world conditions. Additionally, they detail the selection of 18 object categories for annotation based on their relevance and frequency in real-world applications.

An example image taken from DOTA. (a) A typical image in DOTA consisting of many instances from multiple categories. (b), ©, (d), (e) are cropped from the source image. We can see that instances such as small vehicles have arbitrary orientations. There is also a massive scale variation across different instances. Moreover, the instances are not distributed uniformly. The instances are sparse in most areas but crowded in some local areas. Large vehicles and ships have large ARs. (f) and (g) exhibit the size and orientation histograms, respectively, for all instances.

The images used in DOTA-v2.0 are from three distinct image source: Google Earth images, GF-2 and JL-1 (GF&JL) satellite images, and the CycloMedia airborne images. Statistical data, as provided in Table 3, outlines aspects like image area, object area, and foreground ratio. Notably, Google Earth images, thoughtfully selected, constitute the majority of positive samples, although negative samples also play a crucial role in mitigating sample bias. The distributions of objects in GF&JL satellite images and CycloMedia airborne images closely resemble real-world scenarios. It’s worth noting that DOTA-v2.0 includes both RGB and grayscale images, and the images from different sources undergo specific spectral rendering and bit-length optimization processes. These processes ensure consistency in structure and appearance information, making the images suitable for recognition-oriented tasks. While acquisition date is available for images from GF-2, JL-1, and CycloMedia, only 27% of Google Earth images include this information. Given that the primary objective of the task is object recognition in aerial images based on visual cues, the geolocation of images is considered insignificant, and thus, DOTA-v2.0 does not provide geolocation data for its images.

Another valuable meta inforamtion is ground sample distance (GSD), which measures the distance between pixel centers on Earth. GSD is valuable for calculating actual object sizes, which, in turn, can be employed for identifying mislabeled or misclassified instances. Additionally, GSD can be integrated directly into object detectors to enhance the accuracy of category classification for objects with less physical size variation. The authors highlight that GSDs vary across the dataset, with different values for images from GF-2, JL-1, CycloMedia, and Google Earth. Moreover, it’s noted that GSD information is missing in 70% of the images within DOTA-v2.0. However, the absence of GSD data does not significantly impact applications that rely on GSD, as machine learning-based methods can be utilized to estimate it.

A distinct characteristic of DOTA is the diverse orientations of objects in overhead view images. Unlike other object detection tasks, these objects aren’t constrained by gravity, resulting in a wide range of possible angles for object orientation. The authors emphasize that this unique distribution of object angles in DOTA makes it an ideal dataset for research on rotation-invariant feature extraction and oriented object detection.

The aspect ratio (AR) of instances is essential for anchor-based models. DOTA considers two ARs for instances: one based on the original Oriented Bounding Boxes (OBBs) and another based on Horizontal Bounding Boxes (HBBs). The distribution of these two ARs is explored in the dataset. Instances exhibit significant variation in aspect ratio, with many instances having a large aspect ratio.

The number of instances per image varies widely in DOTA, with some images containing up to 1000 instances while others have just one instance. This property is compared to other object detection datasets. The density of instances varies across categories, with some categories having significantly denser instances than others. The authors provide quantitative analysis by measuring the distance between instances within the same category and binning them into three density categories: dense, normal, and sparse. The density is measured by calculating the distance to the closest instance.

The authors also note significant improvements in DOTA from earlier versions (DOTA v1.0 and DOTA v1.5), which included addressing challenges related to tiny objects, large-scale images, and multi-source overhead images. In DOTA-v2.0, there are 18 common categories, 11,268 images, and 1,793,658 instances, with the addition of new categories like airport and helipad. The dataset is divided into train, val, test-dev, and test-challenge (not available at download source - comm. dninja) subsets, each with specific proportions to avoid overfitting. Additionally, two test subsets, test-dev and test-challenge, have been introduced for evaluation, following a similar structure to the MS COCO dataset.

In summary, the authors of the dataset have made significant contributions to the field of object detection in aerial images by providing a comprehensive dataset, baselines, and tools to facilitate research and development in this domain. They have addressed various challenges and limitations to create a more robust benchmark dataset for oriented object detection in aerial images.