Can textboxes++ train data which classes is more than 2(traditional is "background" and "text")?

MhLiao / TextBoxes_plusplus

TextBoxes++: A Single-Shot Oriented Scene Text Detector

Other

954 stars 279 forks source link

Can textboxes++ train data which classes is more than 2(traditional is "background" and "text")? #93

Open zq1335030905 opened 5 years ago

zq1335030905 commented 5 years ago

I want to ask if I want to train my own dataset which have 11 classes , can I train it by textboxes++ ? If I train it by textbox++ , which files should I rectify ? I try to rectify the class nums and some information in labelmap_voc.prototxt and modelConfig.py , but I encounter the error above: Cannot copy param 0 weights from layer 'conv4_3_norm_mbox_conf'; shape mismatch. I don't know how to fix it ? Is there anybody can teach me ? Thanks a lot !!

MhLiao commented 5 years ago

You should also change your network output channel to fit your classes number. the num_classes in modelConfig.py.

zq1335030905 commented 5 years ago

Thank you! I have this problem because of the processing of my dataset . Now I have solve this problem ,thank you very much for your reply~