POINTOUT: Build Datasets for Object Detection on Satellite Imagery.
Deep Learning is revolutionizing the computer vision field in many industries, and now it is the turn for Earth Observation (EO). A lot of interesting applications can take advantage of satellite imagery, such as traffic management, construction site monitoring or environment protection. Working with EO imagery comes with additional challenges when compared with traditional Machine Learning (ML) vision problems. Nevertheless, they all have something in common, and that is the need of labelled datasets. Different datasets exist today for training object detection algorithms on common images (COCO, Pascal VOC) but that is not the case for EO. In this context, the POINTOUT project (developed by Starlab Barcelona S.L and funded by ESA) is intended to provide an open, collaborative web platform where ML/EO professionals, students and enthusiasts can explore and annotate objects directly on a map in order to build datasets for object detection tasks.
Disclaimer: The platform is in its first stages of development, so keep in mind this is an on-going work with limited functionality.
In order to access the platform, please go to targetdetection.com. You will see the landing page of the project with some information and contact options. Please get in touch if you want to know more about the project !
Once you go to the app the first thing that you will see is an interactive world-wide map where you can navigate to your area of interest. In order to explore annotations by other users or start making your own, you will need to register. Then, you will be able to access logging with your username and password.
Select a label and add some annotations
You can select a label from the list on the side panel. To generate new labels, click on the draw toolbox and start drawing boxes around the objects you want to annotate. Once you are happy with your annotations, click on the save button in order to generate an image corresponding to the current viewport along with the annotations.
If you are interested in using POINTOUT to generate a dataset for an object not listed, get in touch with us at email@example.com and we will add it !
Explore and rate annotations by other users
When you select a label, you can see directly on the map annotations made by you (colored background) and other users (transparent background). You can also rate the annotations, which is reflected in the color of the box (red for not rated, yellow for more than one positive rating and green for three or more positive ratings).
You can also download a dataset clicking on the download tab. A filtering option can be used to download only your annotations, restrict the dataset to an area of interest and annotations quality.
Once you download a dataset, you can use it to train your own object detection algorithms using your favorite neural networks framework.
Detect objects on the map
If a model for inference is available for a particular label, the detect tab will be active. Just go to your area of interest, choose a zoom level to perform detection and click on the detect button. Detected objects will be shown in the map.
Training a model directly through the platform
This is an experimental feature only available for demo purposes. If your an interested in see that in action, get in touch with us !
Get in touch !
If you are:
- An ML/EO professional, student or enthusiast interested in providing feedback to improve the platform.
- A business developer interested in a platform like Pointout tailored for your use cases.
Contact us at firstname.lastname@example.org.
Follow us on Twitter @Pointoutproject !
If you are attending the Living Planet Symposium in Milan from May 13th to 17th, please meet us at the poster session on May 16th !