search

clovaai / CRAFT-pytorch

Official implementation of Character Region Awareness for Text Detection (CRAFT)
MIT License
3.13k stars 887 forks source link

Create the bounding box/heatmap for small size symbols #59

Open thaihaquang opened 5 years ago

thaihaquang commented 5 years ago

@ClovaAIAdmin thanks for your nice project. I am finetuning the model with my own dataset including symbols ( . , / \ ~ #). I am wondering how to create the character-level heatmap and affinity heatmap for symbols with small size such as dots. There are 2 ways as below. Do you have any suggestions for me? image Thank you very much!

lamhoangtung commented 5 years ago

@thaihaquang I tried the first way. Not really effective if they spread too far away through :3

YoungminBaek commented 5 years ago

@thaihaquang Sorry for the late reply.

Special characters like periods are hard to deal with. We used the first way by following the character bounding boxes in SynthText dataset. Also, during the process of generating pseudo-GT, the intermediate output from the watershed labeling is the first case because the box covers the region where only the actual characters are located.

However, affinity maps can break links because they aren't able to combine small letters like period or comma.

MrAsifKhan commented 2 years ago

@thaihaquang how is your model performing on the symbols like . , [ ] : ?? because I want to fine tune CRAFT on my dataset which has a combination of Alphabets and symbols. any suggestions are welcomed..