I have spent some time in understanding your paper. I have some questions on domain adaptation
point 1:
After 1st stage of training for several iterations for a good classifier and regressor, you change the dataset to change the labels to 0.5 , freeze the domain mixer/classifier layers , by keeping the lr_mult: =0. train again(stage 2) for several iterations ? Please spare some time to correct me if I am wrong
point 2:
Note that we train the network in a circular way, stage1
is alternate with stage2. This is because domain classifier is
based on current feature extractor. If we modify the extractor
in stage2, then we need to refine new classifier in stage1.
So it is a circular process
After 2 stages of training you have already got the pose regressor without any gap between syn and real data. Then why do we need to do it circularly again? Could you please explain that? " This is because domain classifier is based on current feature extractor" , but feature extractor changes in every iterations anyway? I didn't understand this point ! Thanks in advance .
I have spent some time in understanding your paper. I have some questions on domain adaptation
point 1:
After 1st stage of training for several iterations for a good classifier and regressor, you change the dataset to change the labels to 0.5 , freeze the domain mixer/classifier layers , by keeping the lr_mult: =0. train again(stage 2) for several iterations ? Please spare some time to correct me if I am wrong
point 2:
After 2 stages of training you have already got the pose regressor without any gap between syn and real data. Then why do we need to do it circularly again? Could you please explain that? " This is because domain classifier is based on current feature extractor" , but feature extractor changes in every iterations anyway? I didn't understand this point ! Thanks in advance .