This repoistory contains notebooks for training and running inference on DETR (Detection Transformers) for scene text detection. For directly running the inference code, please reach out to me for a trained model file (trained over 30 epochs on a partial ICDAR dataset which gives descent predictions). A huge thanks to NielsRogge implementation from which most of the DETR code is adapted from.
The main changes for adapting the DETR for text detection are:
- Text detection datasets are typically in ICDAR format which need to be converted to COCO format
- A single "text" category id of 0 (2 in total, one being for the background). So num_classes = 1
This work can also be interesting to people wanting to train DETR on their own custom dataset.