SOVABench
收藏资源简介:
SOVABench是由里程碑系统公司、巴塞罗那大学等机构联合构建的车辆监控行为检索基准数据集,包含9,882条从MEVA和VIRAT监控视频中提取的样本。该数据集创新性地以相反动作对(如开门/关门)为核心构建,包含1,423个查询视频和14类动作标签,视频时长集中在1-10秒,通过空间裁剪突出主体动作。数据集采用双协议评估体系:跨动作对检索(Inter-pair)评估动作区分能力,动作对内检索(Intra-pair)检验时序理解能力,为多模态大模型在监控场景中的行为分析和时序推理提供了标准化测试平台。
SOVABench is a benchmark dataset for vehicle surveillance behavior retrieval jointly constructed by Milestone Systems, University of Barcelona and other institutions. It contains 9,882 samples extracted from MEVA and VIRAT surveillance videos. This dataset innovatively takes opposite action pairs (such as opening/closing a door) as its core construction framework, including 1,423 query videos and 14 categories of action labels. The durations of the videos are mainly concentrated between 1 and 10 seconds, and the main subjects' actions are highlighted through spatial cropping. The dataset adopts a dual-protocol evaluation system: Inter-pair retrieval is used to evaluate action discrimination ability, while Intra-pair retrieval tests temporal understanding ability. It provides a standardized test platform for multimodal large language models (LLMs) to conduct behavior analysis and temporal reasoning in surveillance scenarios.




