PhysicalAI-VANTAGE-Bench-Subset
收藏资源简介:
VANTAGE-BENCH是首个专为评估固定基础设施摄像头捕获视频的视觉理解能力而设计的公开基准数据集。该数据集涵盖三个现实领域——仓库、智能城市/智能交通系统(ITS)和智能空间,涉及六种时空视频理解任务,包括视频问答(VQA)、时间定位、密集视频字幕生成、事件验证、空间定位和时空跟踪。数据集仅用于评估目的。数据收集方法采用混合方式:人类、合成和自动化。视频数据来源包括供应商提供的素材、合成生成以及公开抓取的来源。标注方法同样为混合方式:人类、合成和伪标注。数据集包含312个独特视频,总计42GB存储空间,具体量化数据包括各个任务的视频数量和条目数。数据集格式为视频(mp4)和图像(jpg),目录结构按任务分类,评估通过HuggingFace上的评估服务器进行。
VANTAGE-BENCH is the first publicly available benchmark dataset specifically designed to evaluate the visual understanding capabilities of videos captured by fixed infrastructure cameras. The dataset covers three real-world domains—warehouses, smart cities/Intelligent Transportation Systems (ITS), and smart spaces—and involves six spatiotemporal video understanding tasks, including Video Question Answering (VQA), temporal localization, dense video captioning, event verification, spatial localization, and spatiotemporal tracking. The dataset is intended solely for evaluation purposes. Data collection methods are hybrid: human, synthetic, and automated. Video data sources include vendor-provided footage (GoPro shots of warehouse and smart space environments), synthetic generation (DriveSim collision and multi-camera scenarios), and publicly scraped sources (Dubuque highway/ITS footage). Annotation methods are also hybrid: human, synthetic, and pseudo-labeling. Annotations for VQA, dense video captioning, and temporal localization are primarily human-generated, while spatial localization labels (2D/3D bounding boxes, referring expressions) combine human annotation with pseudo-labeling processes (detection + SAM for spatial pointing), and event verification labels are human-curated. The dataset contains 312 unique videos, totaling 42GB of storage, with specific quantitative data including the number of videos and entries for each task. The dataset format is video (mp4) and images (jpg), with a directory structure organized by task. Evaluation is conducted via an evaluation server on HuggingFace, where prediction results are submitted to compute metrics and update the public leaderboard.




