IntriAdapt is a proposed method for test-time adaptation in metric geometry estimation. The authors point out that recovering metric 3D geometry from images is fundamentally ill-posed because of perspective projection ambiguity. Current geometry models incorporate supplied or predicted intrinsics during inference, but do not use them to supervise.
IntriAdapt changes that by using intrinsics as a guide for adaptation at test time. The goal is to improve generalization when a model is applied to new images, treating intrinsic parameters as a geometric constraint rather than just auxiliary inputs.
Because the abstract is brief, full implementation details and evaluation results are not included here. The paper's contribution is to reframe camera intrinsics as a supervisory signal, which could be a useful direction for making metric geometry estimation more robust across unseen settings.