SOTA端到端算法如何设计？CVPR'25 WOD纯视觉端到端比赛Top3技术分享~

Core Insights - The article discusses the results of the 2025 Waymo Open Dataset End-to-End Driving Challenge, highlighting the advancements in end-to-end autonomous driving systems and the shift towards using large-scale public datasets for training models [2][18]. Group 1: Competition Results - The champion of the competition was the EPFL team, which utilized the DiffusionDrive model, nuPlan data, and an ensembling strategy [1]. - The runner-up was a collaboration between Nvidia and Tubingen teams, which also referenced DiffusionDrive and SmartRefine, employing multiple datasets to demonstrate the importance of training data quality [1][22]. - The third place was secured by Hanyang University from South Korea, which focused on a simplified structure using only front-view input and vehicle state [1][3]. Group 2: Methodology - The UniPlan framework was introduced, leveraging large-scale public driving datasets to enhance generalization in rare long-tail scenarios, achieving competitive results without relying on expensive multimodal large language models [3][18]. - The model architecture is based on DiffusionDrive, which employs a truncated diffusion strategy for efficient and diverse trajectory generation [4][6]. - The diffusion decoder utilizes a cross-attention mechanism to refine trajectory predictions based on scene context [5][6]. Group 3: Data Processing - The nuPlan dataset was processed to create a diverse training set, resulting in 90,000 samples by applying a sliding window approach [7]. - A similar filtering strategy was used for the WOD-E2E dataset, generating 35,000 training samples and 10,000 validation samples [8]. - The model was trained on a powerful computing setup with four H100 GPUs, achieving significant training efficiency [10]. Group 4: Experimental Results - The performance was evaluated using Rater Feedback Score (RFS) and Average Displacement Error (ADE), with various configurations tested [12][17]. - The results indicated that the combined training of WOD-E2E and nuPlan datasets led to slight improvements in average RFS, particularly in long-tail categories [23]. - The analysis showed that while additional datasets generally provide benefits, the quality of the data sources is more critical than quantity [39]. Group 5: Conclusion - The article emphasizes the potential of data-centric approaches in enhancing the robustness of autonomous driving systems, as demonstrated by the competitive results achieved with the UniPlan framework [18][39].