Real-Time monocular depth estimation using synthetic data with domain adaptation via IST

4 Contributions:

1. synthetic depth prediction - a directly supervised model using a light-weight architecture with skip connections that can predict depth based on high-quality synthetic depth training data.
2.  domain adaptation via style transfer - a solution to the issue of domain bias via style transfer
3. efficacy - an efficient and novel approach to monocular depth estimation that produces pixel-perfect depth
4. reproducibility - simple and effective algorithm relying on data that is easily and openly obtained.

Limitations:

The biggest issue is that the approach is incapable of adapting to sudden lighting changes and saturation during style transfer. When the two domains significantly vary in intensity differences between lit areas and shadows(as is the case with our approach), shadows can be recognized as elevated surfaces or foreground objects post style transfer.