HorizonWeaver Editing Dataset
收藏资源简介:
HorizonWeaver数据集由NEC美国实验室与西蒙弗雷泽大学联合构建,包含25.5万张驾驶场景图像,覆盖天气变化、交通流调整等13类编辑任务。该数据集创新性地融合了Boreas、nuScenes和Argoverse2的实景与合成数据,通过自动化配对生成和VLM验证流程确保数据质量,支持多粒度语义编辑研究。其核心价值在于解决自动驾驶领域密集场景下的内容保持与指令对齐难题,显著提升BEV分割等下游任务性能33%。
The HorizonWeaver dataset was jointly constructed by NEC Laboratories America and Simon Fraser University. It comprises 255,000 driving scene images covering 13 types of editing tasks such as weather variation and traffic flow adjustment. This dataset innovatively integrates real and synthetic data from Boreas, nuScenes, and Argoverse2, ensures data quality via automated pairing generation and VLM-based validation workflows, and supports multi-granularity semantic editing research. Its core value lies in resolving the challenges of content preservation and instruction alignment in dense autonomous driving scenarios, and it significantly boosts the performance of downstream tasks such as BEV segmentation by 33%.

- 1HorizonWeaver: Generalizable Multi-Level Semantic Editing for Driving Scenes西蒙弗雷泽大学; NEC美国实验室 · 2026年



