OrientedFormer: An End-to-End Transformer-Based Oriented Object Detector in Remote Sensing Images
OrientedFormer is an end-to-end transformer detector for remote sensing images that handles objects with arbitrary rotations. It uses three modules: Gaussian Positional Encoding to encode rotation, Wasserstein Self-Attention to capture geometric relationships, and Oriented Cross-Attention to align object values with positional queries. This approach successfully resolves challenges related to object orientation, lack of geometric context in attention, and misalignment.