syguan96/DynaBOA

Implements test-time bilevel online adaptation to refine a pre-trained mesh reconstruction model on out-of-domain videos without retraining. Uses a dynamic weighting strategy between inner and outer optimization loops to adapt the model to target domain characteristics like novel camera parameters and motion patterns. Integrates with AlphaPose/OpenPose for 2D keypoint detection and supports real-time inference via webcam or video input.

228 stars. No commits in the last 6 months.