Truncate-Split-Contrast: A Framework for Learning from Mislabeled Videos

Authors: Zixiao Wang, Junwu Weng, Chun Yuan, Jue Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on three well-known benchmark datasets for video classification show that our proposed tru Ncat E-split-contr As T (NEAT) significantly outperforms the existing baselines.
Researcher Affiliation Collaboration Zixiao Wang1*, Junwu Weng2, Chun Yuan1, Jue Wang2 1 Tsinghua University 2 Tencent AI Lab
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access to source code for the methodology described.
Open Datasets Yes We conduct experiments on three large-scale video classification datasets, Kinetics-400 (K400), Mini-kinetics (K200), and Something-Something-V1 (Sth V1).
Dataset Splits Yes K400 This is a large-scale action recognition dataset with 400 categories, and it contains 240k videos for training, and 20k (50 per category) for validation. ... K200 This is a balanced subset of K400 with 200 categories. In this dataset, each category contains 400 videos for training and 25 videos for validation. ... Sth V1 Something-Something is a challenging dataset focusing on temporal reasoning. ... It contains 86k videos for training and 12k for validation.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper mentions
Experiment Setup Yes All training hyper-parameters are adapted from the original work (Lin, Gan, and Han 2019) and detailed in Appendix. Following previous work, we choose τ = 0.1 for all experiments in this paper. A memory bank with the size of 16 K is added for all contrastive learning loss to provide ample training positive-negative pairs.