Hallucinating Optical Flow Features for Video Classification

Authors: Yongyi Tang, Lin Ma, Lianqiang Zhou

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experimental Results and Discussions
Researcher Affiliation Industry Yongyi Tang , Lin Ma and Lianqiang Zhou Tencent AI Lab {yongyi.tang92, forest.linma}@gmail.com, tomcatzhou@tencent.com
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at: https://github.com/ Yongyi Tang92/Mo Net-Features.
Open Datasets Yes We utilize the Kinetics-400 action recognition dataset [Kay et al., 2017] following the public validation split. It contains over 300 thousand of 10-second video clips with 400 different human action labels in total. and The You Tube-8M dataset is a challenging large-scale multi-labels video dataset, which consists of 6 millions of You Tube videos with 3 labels per video on average.
Dataset Splits Yes We utilize the Kinetics-400 action recognition dataset [Kay et al., 2017] following the public validation split. and We follow the similar data split as in [Lin et al., 2018; Tang et al., 2018] that reserves 15% of video for validation, which contains about 100k video clips.
Hardware Specification No The paper mentions general use of GPUs (e.g., "even with GPUs", "takes the advantages of GPUs for computational accelerations") but does not provide specific hardware details such as GPU models, CPU types, or memory specifications.
Software Dependencies No The paper does not provide specific software dependency details, such as library names with version numbers (e.g., PyTorch 1.9, Python 3.8).
Experiment Setup Yes We empirically set the learning rate to 2e 4 and decrease it by 1/10 every 15 epochs with gradient norms clipping to 1.0. We cease the training while the validation accuracy saturates at around 40 epochs.