Why on DAVIS 2016, the model i trained performs much worse than the trained model you provided on your project web page? about unsupervised_detection HOT 5 CLOSED

Hi @RuiFeiHe, thanks for your interest in our project!
You could for sure improve the results training longer. It is also possible to get slightly difference in performance for each run, but differences should be small. Did you report the best score or just the last one?

from unsupervised_detection.

Hello, @antonilo ,thanks for your quick reply！

I did use the best model for testing.
Considering of the training epoch factor, this time, I ran two identical experiments, each trained for 200 epochs.
And I get the following results:
"The Average over the dataset: IoU is 0.5370813250873963 and MAE is 0.11571520393670992
The Average over sequences IoU is 0.5401700757"
"The Average over the dataset: IoU is 0.5221546808647927 and MAE is 0.07263825114729792
The Average over sequences IoU is 0.5348582947857214"

It seems that there still exist noticeable difference between my trained model and your best model.
Do you have any clues for that？

from unsupervised_detection.

hi @RuiFeiHe, could you make sure that the flag train_partition='trainval'?
Did you change any hyperparameter?
If not, could you repeat the experiment and report results?

from unsupervised_detection.

hello, @antonilo , when I first downloaded your files, it seems that the default value of train_partition is 'train'.
After I set it to 'trainval' , the result I get is very close to yours.
Thanks.

from unsupervised_detection.

@RuiFeiHe Hi, Have you reproduced the number on three datasets reported in paper?

from unsupervised_detection.