Prediction looks bad

[Fang-Haoshu]

Hi,

After training the model, I visualize some output of the model and find that the estimated layout looks bad, even on training set. Here are some results. Does anybody successfully train the network and have good results?

[FengyangZhang]

Sorry, as it's an out-of-interest project, I didn't have much time to revise it and make the output better, in fact I stopped when I get reasonable results, you can do some more revision, or apply the recurrent model in the paper to see if you can get better results.

[Fang-Haoshu]

Hi, they are visualization on training set. So I think the network is under-fitting

[Fang-Haoshu]

I made some modifications and now my classification branch is a little bit over-fitted. The accuracy is 0.99 on training set but only 0.7x on testing set. But the keypoint branch works well so my final accuracy is on-par with the paper. How about you? And any suggestions on the classification branch?

[euivmar]

@Fang-Haoshu @kwangHouzz Can you share your code please? (A pull request for example). If you can share also your models weights I would be very grateful!

[DRAhmadFaraz]

@Fang-Haoshu @kwangHouzz Can you share your modified code please?

my email : [email protected]

I would be very grateful!

[jkstyle2]

@Fang-Haoshu @kwangHouzz Can you please share the code please?

My email : [email protected]

it will be very helpful!!