SIMRDWN: Adapting Multiple Object Detection Frameworks for Satellite Imagery Applications
If the myriad challenges of finding small objects in overhead imagery makes you anxious, we invite you to take a deep breath, relax, and simmer down.
Rapid detection of small objects over large areas remains one of the principal drivers of interest in satellite imagery analytics. A number of previous blogs [1, 2, 3, 4, 5, 6, 7, 8] discussed the YOLT algorithm, which modifies YOLO to rapidly analyze images of arbitrary size, and improves performance on small, densely packed objects. YOLO is just one of many advanced object detection frameworks, however, and algorithms such as SSD, Faster R-CNN, and R-FCN merit investigation as well.
To this end we introduce the Satellite Imagery Multiscale Rapid Detection with Windowed Networks (SIMRDWN) framework. SIMRDWN (phonetically: [SIM-er] [doun]) combines the scalable code base of YOLT with the TensorFlow Object Detection API, allowing users to select a vast array of architectures to apply towards bounding box detection of objects in overhead imagery.
Comparison between Faster R-CNN, R-FCN, SSD, and YOLT with a dataset of aircraft, boats, cars, and airports revealed that the YOLT implementation has both the highest mean average precision (mAP = 0.68), and the fastest inference speed (a minimum of 0.44 square kilometers per second).
In this post we illustrate some of the outputs of SIMRDWN for various architectures trained to find vehicles such as aircraft, boats, and cars. Bounding box labels are better suited for vehicles than building footprints; nevertheless, we also explore the performance of SIMRDWN models on building footprint detection on the recently released SpaceNet Off-Nadir Dataset over Atlanta, Georgia.
In future posts we will explore how the SIMRDWN framework helps inform a number of satellite imagery research areas, such as super-resolution. For the time being, we encourage interested parties to inspect the images below, explore the codebase at github.com/cosmiq/simrdwn, or peruse our arXiv paper (to appear in WACV 2019) for further details.