Qcnet->SmartRefine->Donut：Argoverse v2上SOTA的进化之路~

Core Viewpoint - The article discusses advancements in trajectory prediction models, specifically focusing on three papers: "Query-Centric Trajectory Prediction," "SmartRefine," and "DONUT," highlighting their methodologies and improvements over previous models [1]. Summary of Related Papers Query-Centric Trajectory Prediction - Introduces a query-centric scene encoding paradigm that allows the model to learn representations independent of global spatiotemporal coordinates, enabling reuse of past computations without re-normalization [2][3]. - Proposes a two-stage trajectory decoding paradigm, where an anchor-free query generates trajectory proposals, which are then refined using a refiner based on anchor points [2][3]. SmartRefine - Enhances the refinement process in trajectory prediction by introducing adaptive anchor selection and context range acquisition, allowing for more efficient computation by segmenting future trajectory points [28][30]. - Implements anchor-centric context encoding, transforming surrounding context features into the corresponding anchor point's coordinate system to capture more relevant scene information [34]. - Adopts a recurrent and multi-iteration refinement approach, where each trajectory is divided into segments, and each segment undergoes refinement iteratively, improving overall prediction quality [35][37]. DONUT - Builds upon the QCNet architecture, introducing a proposer and refiner module along with an overprediction mechanism to enhance trajectory prediction accuracy [40][41]. - The model segments trajectories into sub-trajectories, predicting future segments based on previous predictions and adjusting reference points for refinement [41][46]. - Achieves state-of-the-art performance in single-agent trajectory prediction on the Argoverse v2 dataset, demonstrating significant improvements over previous models [48].