search

MhLiao / TextBoxes_plusplus

TextBoxes++: A Single-Shot Oriented Scene Text Detector
Other
954 stars 279 forks source link

Caffe modifications #102

Closed sathibault closed 5 years ago

sathibault commented 5 years ago

Can this be run with a standard build of Caffe? What modifications were made?

MhLiao commented 5 years ago

Several extra layers should be included and recompile. Please check the directory “src/caffe/layers/". You can refer to the official document to know how to register new layers.

sathibault commented 5 years ago

I was able to get this working in OpenCV with some minor changes using the pre-trained model. It performs well on the demo image, but I'm unable to make it work on a sample video. The video has cars with clear text on license plates and easy to read signs. The only boxes detected are an apartment window and a car grill.

sathibault commented 5 years ago

I can see now that the text is too small if the image is scaled down (12 pixels high in full resolution). Do you have any sense for the range of text heights that this is able detect?

MhLiao commented 5 years ago
  1. You can try to enlarge the size of the input image. What's your full resolution?
  2. Do you have training data? If yes you can finetune it with your training data.
sathibault commented 5 years ago

It is working fine on the full resolution images 1920x1040. I was just trying to reduce the size to improve performance.