Short, practical explanations for terms used across my projects. (Only the terms I choose to include appear here.)
Domain Adversarial Neural Network: learns features that transfer across domains via adversarial training.
Used when training data comes from one domain (simulation) but deployment is another (real observations).
The model learns representations that are predictive for the task while being hard to distinguish by domain.
Typically implemented using a gradient reversal layer and a domain classifier head.
Goal: reduce performance drop caused by domain shift.
A 3×3 projective transform that maps points between two planes (e.g., camera view → top-down).
Common in multi-camera fusion and bird’s-eye-view (BEV) mapping.
Assumes points lie on (or near) a plane, like a road surface.
Estimated from point correspondences (e.g., manually picked points or calibration targets).
Used to warp images into a shared coordinate frame for stitching or tracking.