News broadcast text localization dataset
The set contains 4225 images, extracted from video news broadcasts. They contain both text added by the news service and naturally occurring text. The images are all of 748*432 size.
The dataset can be downloaded at: https://www.dropbox.com/s/l076cfy18nmgvgc/euronews_frames.7z?dl=0
The provided Matlab code computes precision and recall scores for evaluating ocr text localization performance. The common PASCAL IoU threshold of 0.5 was used. As a demo, we have included a set of results in results_captioncapture.txt
- Download the dataset and provided code
- Modify dataset, annotation and ocr output paths as needed
- Run ocr_eval.m