about the cnn

[Usernamezhx]

thanks for your work. so cool. can you supply more detail about the train way of the cnn. or you use the model trained on the imagenet? thanks

[tanujjain]

Hi, we used a MobileNet pretrained on imagenet dataset. The only modification we did was removing the dense layer and using the last Global Average Pooling layer to generate encodings. Pretty standard thing to do :)

[gaussiangit]

@tanujjain What if we use bigger model or recent more accurate ? Do you think the results would be much accurate ? Like if we want cluster similar images?

[qiankl]

@tanujjain What if we use bigger model or recent more accurate ? Do you think the results would be much accurate ? Like if we want cluster similar images?

@gaussiangit Hello, I wanna cluster similar images too. For normal photos CNN is doing well. Now I encounter a problem that this CNN(mobilenetv3 pretrained on ImageNet) performs bad on mobilephone screen shots/text. I think it maybe because this model never see text samples in ImageNet. Did you have the same problem or do you have some experience on such case? Is there any other model can handle this task? Thanks