I want to create a cnn to draw bounding boxes around individual handwritten words. Ideally I would input a picture of a filled piece of notebook paper and get a cropped image of each word (they don't have to be in order or classified). I would like to use YOLOv2 with a fine grid (maybe 19x38) to obtain the bounding boxes, but I have heard that it does not preform well on small objects. Before I start researching and trying to implement YOLOv3, I wanted to come here to make sure that YOLOv2 would indeed not work for this application and if YOLO would even work at all. Should I use YOLOv3 with only the smallest scale? I was also wondering if I could reduce the number of filters/layers since I don't have any classes and I am only trying to detect words. Should I tweak the initial filter size to detect only small objects? I understand I am asking a lot of questions, I just need some guidance on how to structure the network. Should I use a custom architecture or one of the standardized YOLO ones. I am unfamiliar with how the number of filters/layers and the size of each filter affects what information will be picked up/preserved by the network. Any advice would be greatly appreciated.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.