Aurora11111

@Zhangpeixiang I think it because of your trained picture's size problem ,when you want to detect the bigger size picture you should trained bigger picture, and the test picture's size...

@Zhangpeixiang I used the 360w datasets download from Internet and generate about 100w datasets to simulate the real scene. you can search for 360w datasets on Internet

@Zhangpeixiang maybe you can use some datasets from this project: https://github.com/xiaomaxiao/keras_ocr

@Zhangpeixiang I used the icdar2017 in traget detection. you can down load icdar2017:链接:http://rrc.cvc.uab.es/?ch=7&com=downloads you can reference this：https://github.com/Aurora11111/text-detection-train-yolov3

I think 0.5 is enough for iou.

@Zhangpeixiang reduce your batch size or resized your image size smaller

没有什么影响的吧

me too! have you soved it?

thanks for abc3698, I solved this problem!

是不是我迭代的次数不够，训练的人比较多的话是设置多少迭代次数合适？目前我保证的是每个人有１０分钟的音频