Saicoco

ok, 我试一下，3q；不熟悉前端，尝试把avoid改成auto没成功实现自动分页

Got it @jakeret

@314rated I get the same shape with input by padding large value to input, and you can center crop the output in input_size when you get the output image.

"origin CTPN" means the model release by paper, or caffe repo

ok. get it

opencv 4 can solve it

可以获取最小外界旋转矩形进行仿射变换进行扣取文本

如果图缩放1/4, fps可以到20+

Maybe you can try this`stream.file_obj.read()` to get bytes: ``` def audio_stream_to_tensor_and_meta(element): path, stream = element stream = stream.file_obj.read() ... return audio_tensor, meta ```

Just run as the normal caffe