CRAFT Paper Reading

Detection Structure

• Instead of utilizing Word-level bounding box, CRAFT deals it with in character level.
• Process of Detection are: 1) Localizing individual characters -> 2) Mapping affinity, or closeness, of two adjacent characters-> 3) Post Processing based on region score from (1) and affinity score from (2)

Training Process

• Character Region Score changes as training epoch progresses.
• Region Score represents probability of the given pixel is the center of the character
• Affinity Score represents probability of the space between adjacent characters

Inference(Post Processing)

Experiments results were yielded like the following, according to the trained datasets:

Questions

• What are pseudo ground truths?
• What are synthetic ground truths?

References

• arxiv paper
• Line Conference Video
• Github Repo, although training code is omitted