DiSCO: Differentiable Scan Context With Orientation

Global localization is essential for robot navigation, of which the first step to retrieve a query from map database. This problem called place recognition. In recent years, LiDAR scan based recognition has drawn attention as it robust against appearance change. this letter, we propose LiDAR-based method, named Differentiable Scan Context with Orientation (DiSCO), simultaneously finds at similar and estimates their relative orientation. The orientation can further be used initial value down-stream local optimal metric pose estimation, improving estimation especially when large between current retrieved exists. Our key idea transform feature into frequency domain. We utilize magnitude spectrum descriptor, theoretically rotation-invariant. addition, on differentiable phase correlation, efficiently estimate global using spectrum. With such structural constraints, network learned in an end-to-end manner, backbone fully shared by two tasks, achieving better interpretability lightweight. Finally, DiSCO validated three datasets long-term outdoor conditions, showing performance than compared methods. Codes are released https://github.com/MaverickPeter/DiSCO-pytorch.