Annotation type

Object detection

Machine learning object detection is awesome and focuses on detecting certain objects on images, so it’s utilized in diverse areas, for instance: housing infrastructure, facial detection, self-driving cars, crop field detections, and so on.

Development Seed's Data team has a wide experience working in this field, and for many years the dream has been to build a powerful workflow that allows them to produce high-quality data sets in a short time.

Input, tools, and output diagram for object detection.

The table below indicates the Data team's average on the objects annotation.

Classes per image Images per hour Min/Max image size in pixels
1 120 256x256 / 2048x2048
2 100 256x256 / 2048x2048
3 80 256x256 / 2048x2048

Data team's object detection tools

The list below shows the tools that the Data team uses for image object detection work.


Computer Vision Annotation Tool (CVAT) is a free, open-source, web-based image and video annotation tool. Is easily setup the CVAT server.

It shows the building detection, in this case, is necessary to draw a bbox that cover the building in order to identify to which building belongs and also the building should be classified according to its characteristic for instance: the last building is fully built, the building material is wood polished and it is residential.

Java OpenStreetMap editor

Java OpenStreetMap editor, known as JOSM is an open-source tool. The JOSM tool works with map tiles. JOSM allows drawing polygons easily in large areas and moving around the map tiles. JOSM works together with osm-seed a backed tool for storing the version of each polygon.

The annotation for object detection is fast with JOSM, in this example, the object to annotate is the school, so the bbox needs to cover the school boundary.

Data team's object detection projects