MS-ASL
收藏资源简介:
手语识别是一个具有挑战性且经常被低估的问题,包括在多个流上异步集成的多模态发音体(手形、方向、运动、上身和面部)。在这种情况下学习强大的统计模型需要大量数据,尤其是要应用该领域的最新进展。然而,由于转录这些非书面语言的巨大成本,标记数据是手语的稀缺资源。我们提出了第一个真实的大规模手语数据集,包含超过 25,000 个带注释的视频,我们使用来自符号和相关动作识别的最先进方法对其进行了全面评估。与当前最先进的技术不同,该数据集允许在具有 200 多个签名者的现实环境中研究对看不见的个体(签名者独立测试)的泛化。以前的工作主要处理有限的词汇任务,而在这里,我们涵盖了具有挑战性和不受约束的现实生活记录条件下的 1000 个标志的大类。我们进一步提出从视频分类中知道的 I3D,作为一种强大且合适的手语识别架构,其性能大大优于当前的最先进技术。该数据集对社区公开可用。
Sign language recognition is a challenging and often underestimated problem, involving multimodal articulatory elements (handshape, orientation, movement, upper body posture, and facial expressions) that are asynchronously integrated across multiple streams. Training robust statistical models for this task requires large-scale datasets, especially when leveraging the latest advances in the field. However, labeled data for sign language is a scarce resource due to the exorbitant cost of transcribing these non-written languages. We introduce the first real-world large-scale sign language dataset, which contains over 25,000 annotated video clips, and conduct comprehensive evaluations on it using state-of-the-art methods from sign and related action recognition. Unlike existing state-of-the-art datasets, this dataset enables research on generalization to unseen individuals (signer-independent testing) in real-world settings with over 200 signers. Previous work has mostly focused on limited-vocabulary tasks, whereas here we cover a large vocabulary of 1,000 signs under challenging and unconstrained real-world recording conditions. We further propose I3D, a well-established architecture for video classification, as a powerful and suitable framework for sign language recognition, which significantly outperforms current state-of-the-art methods. This dataset is publicly available to the research community.

- MS-ASL数据集首次发表,由Gallaudet大学、微软研究院和马里兰大学合作开发,旨在为美国手语(ASL)的计算机视觉研究提供标准化的数据资源。
- MS-ASL数据集首次应用于手语识别和翻译研究,显著提升了ASL识别系统的准确性和鲁棒性。
- MS-ASL数据集被广泛应用于多模态学习研究,促进了手语与语音、文本等多模态数据的融合分析。
- MS-ASL数据集的扩展版本发布,增加了更多的手语词汇和多样化的手语表达,进一步丰富了数据集的内容和应用场景。
- 1MS-ASL: A Large-Scale Data Set and Benchmark for Understanding American Sign LanguageMicrosoft Research · 2019年
- 2Sign Language Recognition Using CNN-LSTM NetworkIEEE · 2020年
- 3A Comprehensive Study on Sign Language Recognition Using Deep Learning TechniquesSpringer · 2021年
- 4Sign Language Recognition Using 3D Convolutional Neural NetworksElsevier · 2022年
- 5Transfer Learning for Sign Language Recognition Using MS-ASL DatasetACM · 2021年



