Long-tailed Video Recognition

长尾视频识别

• 批准号: 2766988
• 金额: --
• 依托单位: University of Bristol
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2022
• 资助国家: 英国
• 起止时间: 2022 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2766988
• 关键词: Long; tailed; Video; Recognition

The rise of deep learning (DL) and the availability of publiclarge-scale vision datasets such as ImageNet [1] have led to significant progress inmodern day computer vision. However, knowledge transfer of DL models frompublic datasets to real world usecases faces multiple challenges due to issuessuch as differences in domains, data scarcity etc. Two of the most significantchallenges are :DL models trained on public datasets give suboptimal performance in novel environments or domains. Therefore they either need to be retrained from scratch or fine-tuned for each domain.Re-training DL models from scratch or fine-tuning them in novel domains may still provide sub-optimal and biased performance due to long-tailed nature of some domains.These shortcomings have stirred up strong research motivations and are popularly known as problems of 'domain adaptation' and 'long-tailed recognition' respectively. Even though there have been various works in these fields, most of them concern with image-based problems. However, the above shortcomings appear in video recognition as well and have been scarcely explored.

深度学习（DL）的兴起和公共大规模视觉数据集（如ImageNet[1]）的可用性导致了现代计算机视觉的重大进展。然而，由于领域差异、数据稀缺性等问题，深度学习模型从公共数据集到实际应用的知识转移面临着诸多挑战。两个最重要的挑战是：在公共数据集上训练的深度学习模型在新的环境或领域中表现不佳。因此，他们要么需要从头开始重新训练，要么需要针对每个领域进行微调。由于一些领域的长尾特性，从头开始重新训练DL模型或在新领域对其进行微调仍然可能提供次优和有偏差的性能。这些缺点激起了强烈的研究动机，并分别被普遍称为“领域适应”和“长尾识别”问题。尽管在这些领域已经有了各种各样的研究，但大多数都是关于基于图像的问题。然而，上述缺点在视频识别中也同样存在，并且很少被研究。