Class-Incremental Motion Forecasting

The paper introduces a novel class-incremental motion forecasting framework designed for autonomous vehicles, allowing for the prediction of future trajectories of dynamic agents as new object classes are introduced sequentially. The proposed end-to-end system employs a 3D-to-2D keypoint voting mechanism and a query feature variance-based replay strategy to mitigate catastrophic forgetting while adapting to new classes, achieving strong performance on datasets like nuScenes and Argoverse 2. This approach is significant for practitioners as it enhances the adaptability of motion forecasting models in real-world scenarios where object classes may evolve over time, thus improving the robustness of autonomous navigation systems.