Exploring Temporally Dynamic Data Augmentation for Video Recognition

Authors: Taeoh Kim, Jinhyung Kim, Minho Shim, Sangdoo Yun, Myunggu Kang, Dongyoon Wee, Sangyoun Lee

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To verify the effectiveness of the proposed method, we conduct extensive experiments on the video recognition task, where Dyna Augment reaches a better performance than the static versions of state-of-the-art image augmentation algorithms. The experimental results also demonstrate the generalization ability of Dyna Augment.
Researcher Affiliation Collaboration Taeoh Kim1, Jinhyung Kim2, Minho Shim1, Sangdoo Yun3, Myunggu Kang1, Dongyoon Wee1, Sangyoun Lee4 1NAVER Cloud, AI Tech. 2LG AI Research 3NAVER AI Lab 4Yonsei University
Pseudocode No The paper describes the Fourier Sampling using mathematical equations (Eq. 1) and textual descriptions, but does not provide structured pseudocode or algorithm blocks.
Open Source Code Yes The code will be available at https://github.com/clovaai/dynaaug to ensure reproducibility.
Open Datasets Yes Specifically, we show the effectiveness of Dyna Augment on various video datasets and tasks: large-scale video recognition (Kinetics-400 and Something-Something-v2), small-scale video recognition (UCF101 and HMDB-51), fine-grained video recognition (Diving-48 and Fine Gym), video action segmentation on Breakfast, video action localization on THUMOS 14, and video object detection on MOT17Det.
Dataset Splits Yes Kinetics-400 consists of 240K training and 20K validation videos in 400 action classes. Something-Something-v2 dataset (Goyal et al., 2017) contains 169K training and 25K validation videos with 174 action classes. Both dataset consist of three train/test splits, and we use the first split.
Hardware Specification Yes All our experiments are conducted in a single machine with 8 A100 or 8 V100 GPUs.
Software Dependencies No The paper mentions using "Py Torch (Paszke et al., 2017)" but does not specify a version number for PyTorch or any other key software libraries with version numbers.
Experiment Setup Yes In Table A1, we describe all hyper-parameters and compare our reproduced baselines with the official results in Kinetics-400 (Carreira & Zisserman, 2017) dataset. In Table A2, we describe all hyper-parameters and compare our reproduced baselines with the official results in the other datasets, such as Something-Something-v2 (Goyal et al., 2017), UCF-101 (Soomro et al., 2012), HMDB-51 (Kuehne et al., 2011), Diving-48 (Li et al., 2018) (DV), and Gym288 (Shao et al., 2020) (Gym) models.