Arrow R-CNN for Flowchart Recognition
Bernhard Schäfer, Heiner Stuckenschmidt · 2019
We propose Arrow R-CNN for recognizing the symbols and structure of offline handwritten flowcharts. Arrow R-CNN extends the Faster R-CNN object detection system with an arrow keypoint predictor. This keypoint predictor is used to reconstruct the flowchart structure. We propose a network architecture and data augmentation methods that allow us to train a very deep model on a small publicly available flowchart dataset. Evaluation results show that Arrow R-CNN outperforms existing offline systems by a wide margin. For comparison with existing online flowchart recognizers, we propose an extension for mapping strokes to recognized symbols. Results show that Arrow R-CNN also achieves state of the art in online recognition, even though it does not explicitly leverage stroke information. An ablation study reveals that data augmentation guided by domain knowledge is key to high accuracy on such a small dataset.