Assembly101
收藏资源简介:
Assembly101是一个大规模的多视角视频数据集,专注于理解程序性活动。该数据集由新加坡国立大学创建,包含4321个视频,记录了人们组装和拆卸101种“拆解”玩具车辆的过程。参与者在没有固定指导的情况下工作,序列展示了动作顺序、错误和修正的丰富自然变化。Assembly101是第一个多视角动作数据集,同时包含静态(8个)和自我中心视角(4个)的录制。序列被标注了超过10万个粗略和100万个细粒度动作段,以及1800万个3D手势。我们针对三个动作理解任务进行了基准测试:识别、预期和时间分割。此外,我们还提出了一个检测错误的新任务。独特的录制格式和丰富的标注集使我们能够研究对新玩具的泛化、跨视角转移、长尾分布以及姿态与外观的关系。我们预见Assembly101将作为一个新的挑战,用于研究各种活动理解问题。
Assembly101 is a large-scale multi-view video dataset focused on procedural activity understanding. Developed by the National University of Singapore, it contains 4,321 videos capturing human assembly and disassembly of 101 types of disassemblable toy vehicles. Participants worked without fixed instructions, and the sequences exhibit rich natural variations in action sequences, errors, and corrective actions. Assembly101 is the first multi-view action dataset that incorporates both static (8) and egocentric (4) camera perspectives. The dataset's sequences are annotated with over 100,000 coarse-grained and 1 million fine-grained action segments, as well as 18 million 3D gestures. We conducted benchmark evaluations for three core action understanding tasks: action recognition, action anticipation, and temporal segmentation. Additionally, we propose a novel error detection task. The unique recording format and rich annotation suite enable investigations into generalization to novel toys, cross-view transfer, long-tailed distributions, and the relationship between pose and appearance. We envision Assembly101 as a new challenging benchmark for studying a wide range of activity understanding problems.

- 1Assembly101: A Large-Scale Multi-View Video Dataset for Understanding Procedural Activities新加坡国立大学 · 2022年



