Harness the Sky: Aerial Segmentation Benchmark with DroneDeploy

Stacey Svetlichnaya

TL;DR our latest benchmark in partnership with DroneDeploy trains models for understanding aerial drone imagery—join our efforts here!

The Aerial Segmentation Benchmark

We’re excited to announce our newest benchmark in partnership with DroneDeploy: train machine learning models to segment high-resolution aerial orthomosaics and elevation images from drones. You can read more in the DroneDeploy blog post.

Combine visual & elevation images to identify buildings, ground, vegetation, water, etc.

The challenge

The goal of this project is to collaborate on and advance scene understanding from drone data. Given a high-resolution photograph and elevation of an area mapped by a drone, how well can we locate the different types of objects or landscapes in the scene?  The training data currently has six classes labeled (ground, water, vegetation, cars, clutter, and buildings) and an impressive level of detail at 10cm per pixel. From the benchmark, you can follow two examples (using FastAI or Keras) to reproduce initial baselines on the data and explore many potential improvements: data augmentation and post processing, hyperparameter and architecture tuning, integrating the elevation signal (not currently used), and more.

Long-term impact

Faster and more accurate models for aerial segmentation can help at a massive scale in many domains, including:

At Weights & Biases, we are building tools to support ethical and effective development of machine learning models. Benchmarks give us the opportunity to share datasets and host public collaboration, fostering a transparent environment around a meaningful application.

We are working to support more collaborations and benchmarks in the future, especially on projects relevant to climate change and social good. If you have an interesting dataset or baseline model you’d like to contribute, please reach out at contact@wandb.com.

Join our mailing list to get the latest machine learning updates.