@inproceedings{856c36e6f9f6473eaf365982c5464535,
title = "Self-supervised Compressed Video Action Recognition via Temporal-Consistent Sampling",
abstract = "Compressed video action recognition targets at classifying action class in compressed video, instead of decoded/standard video. It benefits from fast training and inference by reducing the utilization of redundant information. However, off-the-shelf methods still rely on heavy-cost labels for training. In this paper, we propose self-supervised compressed video action recognition method via Momentum contrast (MoCo) and temporal-consistent sampling. We leverage temporal-consistent sampling into MoCo to improve the ability of feature presentation on each input modality of compressed video. Modality-oriented fine-tuning is introduced to applying into the downstream compressed video action recognition. Extensive experiments demonstrate the effectiveness of our method on different datasets with different backbones. Compared to SOTA self-supervised learning methods for decoded videos on HMDB51 dataset, our method achieves the highest accuracy of 57.8\%.",
keywords = "Action recognition, Compressed video, Contrastive learning, Temporal-consistent sampling",
author = "Pan Chen and Shaohui Lin and Yongxiang Zhang and Jiachen Xu and Xin Tan and Lizhuang Ma",
note = "Publisher Copyright: {\textcopyright} 2021, Springer Nature Switzerland AG.; 28th International Conference on Neural Information Processing, ICONIP 2021 ; Conference date: 08-12-2021 Through 12-12-2021",
year = "2021",
doi = "10.1007/978-3-030-92273-3\_20",
language = "英语",
isbn = "9783030922726",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
publisher = "Springer Science and Business Media Deutschland GmbH",
pages = "237--249",
editor = "Teddy Mantoro and Minho Lee and Ayu, \{Media Anugerah\} and Wong, \{Kok Wai\} and Hidayanto, \{Achmad Nizar\}",
booktitle = "Neural Information Processing - 28th International Conference, ICONIP 2021, Proceedings",
address = "德国",
}