There are 0 repository under something-something topic.
Codebase for "Revisiting spatio-temporal layouts for compositional action recognition" (Oral at BMVC 2021).
Something-something-v2 video dataset is splitted into 3 meta-sets, namely, meta-training, meta-validation, meta-test. Overall, dataset includes 100 classes that are divided according to CMU [1] The code also provides a dataloader in order to create episodes considering given n-way k-shot learning task. Videos are converted to the frames under sparse-sampling protocol described in TSN [2]
:floppy_disk: :fire: The fastest and easiest to use TFRecords formatter
Action Detection using 3D CNN