Together with Gemeente De Haag Coders Co. develops PDF2GIS: software that reads old constructions drawings and turns them into a 3D model of the city. We use deep learning to parse different parts of these drawings. One example is text/graphics separation and processing textual information. Existing text detection models do not perform well enough on this type of drawings.
The goal of this project is to create a text detection model that works well with construction drawings. A labeled dataset is available. Possible approaches are:
- Start with one of the pre-trained text detection models, for example EAST and use our labeled dataset and transfer learning to improve the performance of this model on our dataset.
- Start a pre-trained generic object detection model, such as Facebooks RetinaNet and use our labeled dataset and transfer learning to achieve good performance on our dataset.
- Try winning network architectures from the ICDAR’s Robust Reading Competition.