A Pytorch implementation of DB-Text paper

Make awesome things that matter.

Command

Train model

• Modify some configuration in config.yaml

make train

Test model

make test-all

Evaluate model

• For evaluation metric, please refer to MegReader repository

# iou-based Pascal
make ioueval

# overlap-based DetEval
make deteval

History (on TotalText dataset)

Train data

Test data

Test dataset (TotalText)

Heatmap	Polygon	Rotated rectangle

Text-line detection (the model trained on CTW1500 dataset)

Image origin	Text-line detected

Full pipeline

• Recognition model was trained on MJSynth and SynthText dataset

Metric evaluation (DetEval - P/R/HMean)

# for TotalText dataset
make deteval

Method	image size	init lr	b-thresh	p-thresh	unclip ratio	Precision	Recall	F-measure
TotalText-resnet18-fcn (word-level)	640	0.005	0.25	0.50	1.50	0.70	0.64	0.67
CTW1500-resnet18-fcn (line-level)	640	0.005	0.25	0.50	1.50	0.83	0.66	0.74

ToDo

• Support datasets
  • TotalText
  • ICDAR2015
  • SCUT-CTW1500
  • MSRA-TD500
  • COCO-Text
  • Synthtext
  • ArT2019 (included Total-Text, SCUT-CTW1500 and Baidu Curved Scene Text dataset)
• Pytorch-lightning
• Model serving with Torchserve
• Metric callback (P/R/F1)
• IoU-based metric (P/R/F1 - Pascal)
• Overlap-based metric (P/R/F1 - DetEval)
• Model quantization
• Model pruning
• Docker / docker-compose
• ONNX, TensorRT
• Text recognition model

Reference

I got a lot of code from DBNet.pytorch, thanks to @WenmuZhou

• Real-time Scene Text Detection with Differentiable Binarization
• Evaluation metrics
• DBNet.pytorch
• DBNet.keras
• Real-time-Text-Detection
• PSENet.pytorch
• deep-text-recognition-benchmark
• volksdep
• TedEval
• torch2trt