MIMETICS DATASET
The Mimetics dataset contains 713 video clips from YouTube of mimed actions for a subset of 50 classes from the Kinetics400 dataset.
It allows to evaluate on out-of-context human actions methods that have been trained on Kinetics.
Citation
When using or referring to this dataset, please cite our paper [LINK].
@article{mimetics,
author = {Weinzaepfel, Philippe and Rogez, Gr\’egory},
title = {Mimetics: Towards Understanding Human Actions Out of Context},
journal = {arXiv},
title = {2019}}News
- 12/2019: release of version 1.0
Data
You can download the Mimetics dataset.
The zip files contain the following files:
- readme.txt for info about the files
- mimetics_v1.0.csv that contain URLs and timestamps of the YouTube clips following the Kinetics format, you can thus use similar script such as this one to download and trim the videos.
- mimetics_v1.0_clsannot.txt contains annotation about the size of the manipulated objects as well as the relevance of a particular scene for each class. Object sizes are either “no”, “small”, “medium” or “large”. In the paper, we report either for no/small object and medium/large. Scene label is either “yes” or “no”.
- mimetics_v1.0_objectsceneannot.txt contains annotation per video (one video per line) about the presence of a relevant object, of a relevant scene, as well as a binary label whether the actor is a mime artist. The object is a binary label that indicates whether a relevant object is present or not, and similarly for the scene label.