[Paper Review] DOTA: A Large-scale Dataset for Object Detection in Aerial Images

Gui-Song Xia|arXiv (Cornell University)|Nov 28, 2017

Advanced Neural Network Applications40 references137 citations

TL;DR

Introduces DOTA, a large-scale aerial-image object-detection dataset with 15 oriented categories, 188,282 instances, and 2806 high-resolution images; provides baselines for horizontal and oriented bounding box detection.

ABSTRACT

Example datasets of the Interactive Feature Localization in Deep neural networks (IFeaLiD) tool. <strong>Cityscapes</strong> These datasets are based on the image <code>bielefeld_000000_007186_leftImg8bit.png</code> of the Cityscapes dataset. The datasets can be explored online in IFeaLiD: conv2_x (<code>bielefeld_000000_007186_leftImg8bit.png.C1.npz.8.zip</code>) conv3_x (<code>bielefeld_000000_007186_leftImg8bit.png.C2.npz.8.zip</code>) conv4_x (<code>bielefeld_000000_007186_leftImg8bit.png.C3.npz.8.zip</code>) <strong>COCO</strong> These datasets are based on the image <code>000000015746.jpg</code> of the COCO dataset. The datasets can be explored online in IFeaLiD: conv2_x (<code>000000015746.jpg.C1.npz.8.zip</code>) conv3_x (<code>000000015746.jpg.C2.npz.8.zip</code>) conv4_x (<code>000000015746.jpg.C3.npz.8.zip</code>) <strong>DIV2K</strong> These datasets are based on the image <code>0804.png</code> of the DIV2K dataset. The datasets can be explored online in IFeaLiD: conv2_x (<code>0804.png.C1.npz.8.zip</code>) conv3_x (<code>0804.png.C2.npz.8.zip</code>) conv4_x (<code>0804.png.C3.npz.8.zip</code>) <strong>DOTA</strong> These datasets are based on the image <code>P0034.png</code> of the DOTA dataset. The datasets can be explored online in IFeaLiD: conv2_x (<code>P0034.png.C1.npz.8.zip</code>) conv3_x (<code>P0034.png.C2.npz.8.zip</code>) conv4_x (<code>P0034.png.C3.npz.8.zip</code>)

Motivation & Objective

• Motivate and enable robust object detection in aerial imagery by providing a large, diverse, high-resolution dataset reflecting real-world Earth Vision challenges.
• Define oriented bounding box annotations to better enclose objects in aerial scenes and reduce inter-object overlap.
• Establish baselines using state-of-the-art detectors for both horizontal bounding boxes (HBB) and oriented bounding boxes (OBB).
• Highlight challenges in dense scenes, extreme aspect ratios, and large intra-class size variation to guide future algorithm development.

Proposed method

• Collect 2806 aerial images from multiple sensors/platforms with resolutions up to 4000x4000.
• Annotate 188,282 object instances with arbitrarily oriented quadrilateral bounding boxes across 15 categories.
• Provide spatial-resolution metadata for each image to contextualize object sizes.
• Evaluate baselines by adapting Faster R-CNN-type frameworks for OBB and standard detectors for HBB, including patch-based processing due to large image sizes.
• Cropping strategy: generate 1024x1024 patches with stride 512 for CNN processing and merge results with non-maximum suppression.

Experimental results

Research questions

• RQ1 How well do existing object detectors generalize to aerial imagery with highly variable object scales and orientations?
• RQ2 What is the performance gap between detectors trained on horizontal bounding boxes versus oriented bounding boxes on aerial data?
• RQ3 Can oriented bounding boxes provide substantial gains over axis-aligned boxes for densely packed aerial scenes?
• RQ4 How does DOTA challenge detectors in terms of small/large object balance, crowding, and extreme aspect ratios?

Key findings

• DOTA contains 188,282 instances across 15 categories in 2806 high-resolution images, making it the largest annotated aerial-object dataset to date.
• Detectors trained/adapted for oriented bounding boxes (OBB) show notable improvements over axis-aligned bounding boxes (HBB) in crowded and elongated objects.
• Baseline results reveal significant difficulty for small and densely packed objects (e.g., small/large vehicles, ships) in aerial scenes, with mixed performance across categories.
• Cross-dataset experiments show substantial generalization gaps between UCAS-AOD and DOTA, indicating DOTA's greater diversity and challenge.
• Dataset-specific annotation with oriented quadrilaterals provides tighter enclosures and better separation of nearby instances compared to horizontal boxes.