search

xtudbxk / DSRG-tensorflow

a tensorflow version for DSRG (Weakly-Supervised Semantic Segmentation Network with Deep Seeded Region Growing)
75 stars 20 forks source link

Retrain to boost performance #15

Closed lianxxx closed 5 years ago

lianxxx commented 5 years ago

Hi,

Have you tried the retrain step, which can boost the performance by around 1.5%, as mentioned in the original DSRG paper?

I tried to implement the retrain step based on your code, but I got worse results compared to those without retrain. Sad..............................................................................................................

xtudbxk commented 5 years ago

The retrain step indeed improves the performance in my experiment. Maybe you can predict all the labels and use codes from the project aims to implement the origin deeplabv2.

lianxxx commented 5 years ago

Was your retrain model initialized with the weights from the pre-training with DSRG? Could you please kindly provide your re-training hype-parameters: init_learning_rate, learning rate policy, batch_size, number of training epoches? Btw, how much improvements did you get from retraining? Looking forward to your reply. Thank you very much!!!

xtudbxk commented 5 years ago

Yes, it was. And it seems I didn't change the hyper-parameters for the retrain step. And in my experiments which are not same with DSRG, the retrain steps seems to improve about 1% in miou.