EfficientDet For Semantic Segmentation and EAST

Where did I copy it from?

1. Yet-Another-EfficientDet-Pytorch by Zylo117 contains most of the base code to create the EfficientDet, however I cleaned it a little to have only what I need and used lukemela's implementation of EfficientNet directly
2. EfficientDet original paper
3. EAST: An Efficient and Accurate Scene Text Detector inspiration to this repo

What can I do?

1. Object detection - you can still use EfficientDet for object detection, however we recommend sticking to Zylo's code
2. Segmentation - you can train EfficientDet to perform segmentation
3. Text Detection - you can use EfficientDet to segment and predict bounding boxes as proposed on EAST: An Efficient and Accurate Scene Text Detector

Limitations

1. We do not have support for EfficientDet-D7x

Results

1. Our first public result is on ICDAR 2019 Robust Reading Challenge on Scanned Receipts OCR and Information Extraction

   • We obtained an Hmean of 93.76% on test set
   • Training involved only a EfficientNet backbone initialized on ImageNet that was frozen. We trained the rest of the model from a random start.

2. The checkpoint for segmentation on ICDAR2019 Robust Reading Challenge on Arbitrary-Shaped Text can be found here. The experiment containing training code and hyper parameters can be found here

3. Training for Text Detection on DocVQA can be found here and experimento logging here

4. Second result on ICDAR 2019 Robust Reading Challenge on Scanned Receipts OCR and Information Extraction

   • Hmean of 94.34%
   • Pre-training on
     • Segmentation on ArT dataset
     • Text Detection on DocVQA
   • Model checkpoint here
   • Neptune experiment here

TODO

• Add other coeficients for Segmentation and EAST
• Add D7 to Segmentation and EAST
• Add EfficientDet-D7x
• Release model checkpoint for the SROIE 2019 submission
• Add results and model checkpoint for the ICDAR2019 Robust Reading Challenge on Arbitrary-Shaped Text
• Add documentation
• Add testing