About weakly-supervised learning???

[lloydnguyen96]

When you train your CRAFT model in a weakly-supervised manner:

• What are your transformations in transforms.Compose, which transformation and in which order?
• Did you move your batch of preprocessing images from CPU to your dedicated GPU or did you move your CRAFT model from GPU back to CPU for performing preprocessing. How do batch of images and model data flow in your 4 GPU and 1 CPU?
• Did you crop and embed all word boxes from original image to a single new image to create pseudo GT region_maps (same in paper) or did you crop and embed each word box from original image to a separated new image?
• When in weakly-supervised training, you generate char boxes using watershed labeling. So in inference, which way do you use to generate char boxes (not word boxes or polygon)? Thank you very much!!!

[dlmade]

hey if possible can you share a script to achieve character level boxes.