Depth for pseudo-labeled images with privacy blurred faces

[ZoengMingWong]

Hi, thanks for your awesome works!
I have the problem that, for some datasets, like SA-1B and P3M-10K, with blurred faces for privacy consideration, a discontinuous depth map would be generated on the face blurry mask. An example likes this:

How you deal with these samples while training with pseudo labels? Ignoring losses of these regions with masking? Or just no any special processing just like other samples?

Thanks for your response!