Gareth Halfacree

Piero Toffanin, a developer on the OpenDroneMap project, has released a resource-light library for object detection — targeting automated analysis of aerial imagery captured by drones, aircraft, or satellites: GeoDeep.

"Today I’ve released the first version of GeoDeep, a new lightweight library for performing object detection on GeoTIFFs," Toffanin explains of the project, which came about while experimenting with a computer vision plugin for the OpenDroneMap project. "While the primary use case is for use in WebODM [in-browser OpenDroneMap], it works as a standalone library and I hope it can be useful for other use cases outside of WebODM."

GeoDeep is a lightweight Python tool for object recognition in drone mapping imagery, seen here locating cars. (????: Piero Toffanin)

The Python library, which can run on the CPU or use a GPU for acceleration, is designed for use with geospatial rasters (GeoTIFFs), and comes with support for two main models: one for detecting cars based on YOLOv7, and another for detecting tree crowns using the DeepForest Retinanet model. Two other models are also available, but described as "experimental": an updated tree crown detection model using YOLOv9, and a DeepForest Retinanet model for detecting birds.

"The basic idea is to create an ONNX model," Toffanin explains of the process of getting additional models into the library, "and possibly make some modifications to GeoDeep to handle different conventions in model architectures via conditional checking."

More information is available on the OpenDroneMap forum, while the GeoDeep library can be found on GitHub under the reciprocal GNU Affero General Public License 3.

Gareth Halfacree

Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.