Introducing DroneDeploy's Aerial Segmentation Benchmark

October 25, 2019

At DroneDeploy, our mission is to make the sky productive and accessible to everyone. Our company grew out of the complexity of using drones to protect endangered wildlife in South Africa in 2013. Now our customers orchestrate over 1 million drones flights every year in over 180 countries.

Today, we are excited to release the DroneDeploy Segmentation Benchmark challenge. We are partnering with Weights & Biases to run a benchmark to evaluate and encourage state-of-the-art machine learning on aerial drone data. The evaluation environment and leaderboard are available now and include a baseline model to get started.

Our product is being used to capture, collect, analyze, and share imagery in a way that helps enable strategic decision-making and take better actions. In agriculture, this can mean understanding where crops are under stress and adjusting the nutrients for precision farming. On a construction site, this can mean daily site progress monitoring through the ability to rewind and replay the construction process day-by-day to visually inspect the site and verify work done.


We are also proud to see DroneDeploy used in humanitarian aid and disaster relief efforts. Understanding every pixel in an image can go a long way in helping take what was once difficult, time-consuming, or even dangerous, and transform them into simple, cost-effective, automated tasks.

We have built our own proprietary photogrammetry pipelines (called MapEngine) that runs in the cloud on a very large scale. These pipelines process all the imagery datasets as they are collected by the drones around the world. These pipelines handle hundreds of millions of images a year to create pointclouds and 3D models of each site. In many cases, accuracy and speed are mission-critical, particularly to disaster recovery and first response where it can be used to save lives.

Photogrammetry is a computationally expensive task that involves both structure-from-motion to estimate the locations of the cameras, followed by multi-view stereo to create pointclouds. Having a better semantic understanding of the scene being reconstructed can yield much faster and more accurate results. For example, vegetation can be reconstructed in a different way from manmade structures.


Sharing results and insights is key with multiple people working on the same machine learning problems, and having all results consolidated into one tool speeds up engineering and eliminates unnecessary and costly duplicate work. Weights & Biases was easy to set up without any changes to our workflow, and we could immediately centralize all our machine learning training metrics into one place.

The Dataset

We’ve collected a dataset of aerial orthomosaics and elevation images. These have been annotated into 6 different classes: Ground, Water, Vegetation, Cars, Clutter, and Buildings. The resolution of the images is approximately 10cm per pixel which gives them a great level of detail. We’re looking forward to making more data available and encourage more research into the impact this imagery can have in furthering safety, conservation, and efficiency.


We are committed to the responsible and ethical use of machine learning and robotics. Machine learning enhances our ability to understand the meaning of content at scale and we strive to use this information to make the world a better place.

About the Author

Getting started is simple. Request a demo today.