Hi all,
I’m doing a project where I have lots of easily generated synthetic data with labels, and also lots of unlabeled real data, and I’m trying to learn about the best way to leverage the unlabeled data.
There appear to be lots of possible directions to go, but I’m a bit overwhelmed by the possibilities, and there doesn’t seem to be a standard “go to” method ( that I could find. )
From what I’ve seen, there are a few options:
Domain Adaption
- Use adversarial training to create robust features useful in both domains ( DANN, etc )
Unsupervised pre-training
- A few recent papers ( PIRL, Moco, CTC ) use unsupervised pre-training to create robust features for later use in the desired task
Image-to-image translation
- e.g. use CycleGAN or similar to translate images from the labeled domain to unlabeled domain
( I tried this with limited success on my particular data )
My task is keypoint prediction via dense heatmaps, and it seems most of the papers deal with classification, even though the ideas should carry over in some of the methods.
So does anyone have a “go to” for this situation? Would I be better off spending the time just labeling my real data instead of researching these methods?
Thanks a lot
