search

charigyang / itsabouttime

Code repository for "It's About Time: Analog clock Reading in the Wild"
MIT License
70 stars 10 forks source link

Some question about train STN #4

Open xywlpo opened 2 years ago

xywlpo commented 2 years ago

Hi, thank you for your amazing work. I am inspired by your work about alignment the clock. I have tried to use STN for alignment my pointer meter as followings. I picked front meter of each category as a standard meter. And online random homography matrix was generated by code to apply on these standard meters for training STN. My batchsize is 32, and iteration about 120000, trian data is 10000, but the results is not good, and the loss is about 0.5 finally. Could you give my some advice about this task?

Thanks very much. circle29 circle48 circle129

charigyang commented 2 years ago

Hi thanks for your interest, we supervise the STN using synthetic data (i.e. when we generate them we also have GT parameters for the STN). We find that that helps. I suppose if you use our pretrained model for STN it might be not too bad.

xywlpo commented 2 years ago

Hi thanks for your interest, we supervise the STN using synthetic data (i.e. when we generate them we also have GT parameters for the STN). We find that that helps. I suppose if you use our pretrained model for STN it might be not too bad.

Thanks for your reply. I use your pretrained model full+++_st.pth as my initial model weights. If I just use synthetic data to train the model, do you think that will be fine? Could you tell me your loss value when the model training finished? Did you just use synthetic data to train your STN without any real data ?

charigyang commented 2 years ago

STN is trained with synthetic as supervision (though the gradient should flow end-to-end when training recognition). I think we trained synthetic one fine though. Loss ends around 7*10^-3 (for STN) and 0.5 (for recognition).

xywlpo commented 2 years ago

flow

Thanks very much. I will try to generate synthetic data to simulate my real meter, and then to train my model.

sajsxj commented 11 months ago

flow

Thanks very much. I will try to generate synthetic data to simulate my real meter, and then to train my model.

hi, what was the result and could u share ur approach? thank u very much!