Fine-tuning CRAFT

[mazatov]

Hello @ahmed-nady , thanks for the very interesting paper and the presentation. I had a question regarding your section 3.1.1. You mention that you take pre-trained CRAFT and retrain it on your dataset. However, in the same paragraph you mention that CRAFT doesn't have training code. So I was wondering, how do you fine-tune CRAFT for jersey numbers?

In our implementation, the weights of CRAFT detector are initialized by the use of the general pre-trained
model and then is trained with the first subset of S2JN dataset to take into account the distortion of the number 
printed on the player’s shirt. The first subset is splitted into training set containing 1274 player images, 
validation set having 317 player images and the remaining 281 player images are used for testing. 
Because of the lack of CRAFT (Baek et al., 2019b) training code, we supervised the training by providing the
annotations for each digit in jersey numbers.

[ahmed-nady]

Thanks @mazatov. For your question, "Because of the lack of CRAFT (Baek et al., 2019b) training code, we supervised the training by providing the annotations for each digit in jersey numbers". As you know CRAFT (Character Region Awareness for Text Detection) is character based text detection. That means that the annotations are not only bounding box around each a word, but also bounding box around each character in a word. The training of CRAFT is done in a weekly supervised by getting bbox around each char using bbox around a word and that part of implementation is not provided. So, the only way to use/finetune this method is to implement this part or as i did determine bbox around each digit.