search

clovaai / deep-text-recognition-benchmark

Text recognition (optical character recognition) with deep learning methods, ICCV 2019
Apache License 2.0
3.71k stars 1.09k forks source link

Same Prediction for all Items within one Batch during Finetuning #273

Open PascalHbr opened 3 years ago

PascalHbr commented 3 years ago

I am trying to fine-tune the model on single characters (my own dataset). I created the dataset as described within the README, but unfortunately the prediction for each item within a batch is always the same. I mean literally the same - the confidence score for each image is identical, and the network converges towards the most common character. I already tried to change the learning rate. Training a model by scratch results in the same behavior. Does anyone know, why this is the case? Do I have to preprocess my images or anything? Thanks a lot!

arshv27 commented 3 years ago

Hey Please check the following:-

  1. Are you cropping your images to ensure single words are being passed to the model?
  2. Check the resolution of your images. I'm facing a similar issue when evaluating on very low resolution images