Discovering Informative and Robust Positives for Video Domain Adaptation

Authors: Chang Liu, Kunpeng Li, Michael Stopa, Jun Amano, Yun Fu

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4 EXPERIMENTWe conducted experiments to assess our method on two prevalent benchmark datasets for video domain adaptation, specifically UCF HMDB (Chen et al., 2019a) and Epic-Kitchens (Munro & Damen, 2020).
Researcher Affiliation Collaboration 1Northeastern University, Boston, MA, USA 2Konica Minolta, San Mateo, CA, USA
Pseudocode Yes Algorithm 1 Robust Cross-domain Positives for video DA
Open Source Code No The paper states 'all related publications and source codes are cited appropriately' but does not explicitly state that *their* code is being released or provide a link to it.
Open Datasets Yes UCF HMDB is first assembled by Chen et al. (Chen et9al., 2019a) for studying video domain adaptation problem. This dataset is a subset of the UCF (Soomro et al., 2012) and HMDB datasets (Kuehne et al., 2011)... Epic-Kitchens... from the full Epic-Kitchens dataset (Damen et al., 2018).
Dataset Splits No To ensure consistency with prior research, we utilized the training and testing partitions provided by the respective authors in (Chen et al., 2019a; Munro & Damen, 2020). The paper explicitly mentions training and testing partitions, but does not specify validation splits.
Hardware Specification Yes We use four 12G NVIDIA GPUs for training.
Software Dependencies No The paper mentions using I3D and SGD but does not provide specific version numbers for software dependencies like PyTorch, Python, or CUDA.
Experiment Setup Yes We follow the standard pre-train then adapt procedure... train the model... for 3 epochs as a warm start... train models... for 40 epochs in total. For the hyperparameters, we select λ from {5, 10, 15, 20}, α0 from {0.1, 0.25, 0.5, 0.75} and the number of neighbors k from {1, 2, 5, 10, 20}. We set k = 5, λ = 15 and α0 = 0.25 for all datasets. We train all the models end-to-end using SGD with a momentum of 0.9 and a weight decay of 1e-7. We use an initial learning rate of 0.01 for the I3D with a cosine learning rate scheduler for our experiments. We use a batch size of 32 equally split over the two domains.