2.7K videos have steps' class and temporal boundary annotations these videos are used for the Step Recognition and Step Forecasting tasks. CrossTask has 4.7K instructional videos annotated with task name for each video spanning 83 tasks with 105 unique steps. The CrossTask dataset is the other dataset that we used for downstream tasks. Please follow the instructions provided by the authors of the COIN dataset to download the raw videos from YouTube, and their annotations in JSON format.Īfter downloading the dataset, please populate the coin_video_dir and coin_annoataion_json fields in the config files of the downstream evaluation settings accordingly when you need to start a downstream evaluation setting (described below). The average number of steps per video is 3.9. COIN contains 11K instructional videos covering 778 individual steps from 180 tasks in various domains. The COIN dataset is one of the datasets that we used for downstream tasks. Particularly, for pretraining on the subset of HowTo100M with 85K videos that we mentioned in our paper, we used train.csv provided by authors of TimeSformer to obtain the list of video IDs. Please follow instructions from dataset provider to download videos, metadata, and subtitles if needed. HowTo100M is a large-scale unlabeled instructional video corpus. Note that we only used the step 'headline'. The length of each sublist corresponds to the number of steps of that task article each element in the sublist is a dictionary with detailed step information. The wikiHow dataset stored in step_label_text.json is a list with a length of 1053, i.e., 1053 sublists representing 1053 task articles. Please follow the instructions provided here to obtain the data (i.e., step_label_text.json). These wikiHow tasks have at least 100 video samples in the HowTo100M dataset. The version of the wikiHow dataset that we used for both of the Procedural Knowledge Graph construction and model pre-training has a total of 10, 588 step headlines from 1, 053 task articles. WikiHow is a text-based procedural knowledge database. Pip install -r requirements.txt Dataset Preparation wikiHow
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |