search

huizhang0110 / AON

Implementation for CVPR 2018 text recognition Paper by Tensorflow: "AON: Towards Arbitrarily-Oriented Text Recognition"
MIT License
108 stars 32 forks source link

关于bilistm提取特征再reverse? #1

Open sheirving opened 6 years ago

sheirving commented 6 years ago

您好！想问下在共享卷积层提取特征后，采用BiLstm将特征序列化，水平和垂直特征分别再reverse出不同方向的特征，有个疑问是BiLstm本身具有双向性，bw为fw的reverse，这里的再reverse是否有些多余？谢谢！

dabenzhong-wang commented 6 years ago

个人理解，是为了得到ABC和CBA两种输出，解决对180°翻转图片的识别。

MnmohanSinghThakur commented 5 years ago

Please share how to create_tfrecord & supply input image fro demo.py? Also, these files are missing '/share/zhui/mnt/demo_20.tfrecord' from input_data.py

Anybody solved this issue?

gzhcv commented 5 years ago

@sheirving bilstm与特征reverse不是一回事。bilstm是考虑了上文和下文的信息，最重要的还是利用当前帧的(设为t)Input信息x_t来做预测；而特征reverse后，当前帧的Input信息其实来自于xt和x{L-t}(设特征序列长度为L)。