DFNet: Enhance Absolute Pose Regression with Direct Feature Matching

We introduce a camera relocalization pipeline that combines absolute pose regression (APR) and direct feature matching. By incorporating exposure-adaptive novel view synthesis, our method successfully addresses photometric distortions in outdoor environments existing photometric-based methods fail to handle. With domain-invariant matching, solution improves accuracy using semi-supervised learning on unlabeled data. In particular, the consists of two components: Novel View Synthesizer DFNet. The former synthesizes views compensating for changes exposure latter regresses poses extracts robust features close domain gap between real images synthetic ones. Furthermore, we an online data generation scheme. show these approaches effectively enhance estimation both indoor scenes. Hence, achieves state-of-the-art by outperforming single-image APR as much 56%, comparable 3D structure-based methods. (The code is available https://code.active.vision .)