- Synth800k (train only)
- ICDAR 2015
- ICDAR 2013
- ICDAR 2017 MLT
We use model trained on ImageNet dataset as our pre-trained model. The training process includes two steps: first we use Synth800k dataset to train the network for 10 epochs, and then real data is adopted to fine-tune the model until convergence. ... Some blurred text regions in ICDAR 2015 and ICDAR 2017 MLT datasets are labeled as “DO NOT CARE”, and we ignore them in training.
First, longer sides of images are resized from 640 pixels to 2560 pixels. Next, images are rotated in range [−10◦ , 10◦ ] randomly. Then, the heights of images are rescaled with ratio from 0.8 to 1.2 while their widths keep unchanged. Finally, 640×640 random samples are cropped from the transformed images.
Employ shrunk version of the original text regions: 0.3 from each corner according to EAST while the area between the bounding box and the shrunk version is considered as “NOT CARE”.
For each image, 512 hard negative samples, 512 random negative samples and all positive samples are selected for classification. As a result, positive-to-negative ratio is increased from 1:60 to 1:3. And for bounding box regression, we select 128 hard positive samples and 128 random positive samples from each image for training gives 2%.
For loss info see fomulas 1, 2 and 3 in the paper.
Final detection results are produced by applying thresholding and NMS to these positive samples.
Or locality aware NMS as in EAST?
| Precision | Recall | F-measure | |
|---|---|---|---|
| ICDAR 2015 | 88.84 | 82.04 | 85.31 |
| ICDAR 2017 MLT | 79.48 | 57.45 | 66.69 |
| ICDAR 2013 | 86.96/87.32 |
It is expected that the provided for training folder contains unzipped ch4_training_images and ch4_training_localization_transcription_gt from Task 4.4: End to End (2015 edition).
The best result "recall": 0.7746750120365913, "precision": 0.8085427135678392, "hmean": 0.7912466191295796 for ICDAR 2015 was reached at this commit https://github.com/Wovchena/FOTSBasedTextDetection/pull/1/commits/90db90cb915d7d611a7da4380b2cca3add6b9d8b with treshold 0.95 for classification. The pretrained model is here: https://drive.google.com/open?id=1xaVshLRrMEkb9LA46IJAZhlapQr3vyY2