# You Only Look Twice: Rapid Multi-Scale Object Detection In Satellite Imagery

**Authors:** Adam Van Etten  
**arXiv:** [1805.09512](https://arxiv.org/abs/1805.09512)  
**Date:** 2018-05-24  
**Summary page:** https://www.summarizepaper.com/en/paper/1805.09512v1/

## Key points

- Detection of small objects in satellite imagery is challenging
- Deep learning approaches have advanced ground-based object detection
- Applying these techniques to overhead imagery is not straightforward due to large number of pixels and geographic extent per image
- Objects of interest in satellite imagery are often minuscule, making traditional computer vision techniques ineffective
- Authors propose a pipeline called You Only Look Twice (YOLT) for rapidly evaluating satellite images of any size
- YOLT can detect objects of various scales using relatively little training data across multiple sensors
- YOLT achieves high scores for vehicle localization with an F1 score greater than 0.8 on large test images at their native resolution
- YOLT can accurately localize objects as small as 5 pixels with high confidence
- YOLT offers a solution for efficiently detecting small objects in satellite imagery by leveraging deep learning techniques
- Code for implementing YOLT is available on GitHub for further exploration and utilization

## Summary

The detection of small objects in satellite imagery is a challenging problem in the field of satellite imagery analytics. While ground-based object detection has seen advancements with deep learning approaches, applying these techniques to overhead imagery is not straightforward. This is due to the large number of pixels and geographic extent per image, with a single DigitalGlobe satellite image covering more than 64 km2 and containing over 250 million pixels. Additionally, the objects of interest in satellite imagery are often minuscule, spanning only around 10 pixels, which makes traditional computer vision techniques ineffective. To address these challenges, the authors propose a pipeline called You Only Look Twice (YOLT), which can rapidly evaluate satellite images of any size at a rate exceeding 0.5 km2/s. The YOLT pipeline is capable of detecting objects of various scales using relatively little training data across multiple sensors. The authors evaluate the performance of YOLT on large test images at their native resolution and achieve high scores for vehicle localization, with an F1 score greater than 0.8. Furthermore, the authors explore the resolution and object size requirements by systematically testing the YOLT pipeline at decreasing resolutions. They find that even objects as small as 5 pixels can be accurately localized with high confidence using this approach. In conclusion, the proposed YOLT pipeline offers a solution for efficiently detecting small objects in satellite imagery by leveraging deep learning techniques. Its ability to handle large swaths of imagery and detect objects at different scales makes it a valuable tool for satellite imagery analytics applications. The code for implementing YOLT is available on GitHub for further exploration and utilization.