SURFZJY
SURFZJY
Line 32, change 'shufflenetv2' to 'resnet50', you can find more backbone in this file: https://github.com/SURFZJY/Real-time-Text-Detection-DBNet/blob/master/models/model.py
The sufflenetv2 model trianed on ic13 and15 has been released, but for some implementation details are different, the results are not as good as the results in the paper. I...
It can work on CPU, by setting the gpu_id=None and uncommenting torch.cuda.synchronize(self.device).
yes, i am working hard to improve the performance. I will release a better version some days later.
both ok, you can have a try ^_^
If you want to train polygon type dataset, you need to modify the dataloader part.
You can use ICDAR2015 dataset. The genText.py can be used to generate the txt file for train/test in our framework.
> how can we download ICDAR2015 dataset. https://rrc.cvc.uab.es/?ch=4
Maybe your geo_map_cython_lib was not compiled successfully, you can uncomment the line 598-608 and 615-616 in icdar.py and do not use the geo_map_cython_lib .
> Is geo_map_cython_lib not necessary? It can accelerate the training process over 10x. It not necessary.