This repository contains the dataset and code for our ACL 2019 paper:

Identifying Visible Actions in Lifestyle Vlogs

Given a video and its transcript, which human actions are visible in the video?

We provide a Google Drive folder with the raw miniclips.

A miniclip is a short video clip (maximum 1 min) extracted from a YouTube video. We segment the videos into miniclips in order to ease the annotation process. For more details on how the segmentation is performed, see section 3.1 in our paper.

The video youtube urls from which the miniclips were created can be found here, together with the video transcripts and titles.

Important: You can automatically download your own videos, transcripts and perform the text and movement filtering by running the code described in the Youtube processing README section. If you still need the raw videos, e-mail me.

The annotations of the miniclips are available at data/miniclip_action.json. The JSON file contains a dictionary: the keys represent the miniclips (e.g. "4p1_3mini_5.mp4") and the values represent the (action, label) pairs.

The miniclip name is formed by concatenating its YouTube channel, playlist, video and miniclip index. For miniclip "4p1_3mini_5.mp4":

• 4 = channel index
• p1 = playlist index (0 or 1) in the channel
• 3 = video index in the playlist
• mini_5 = miniclip index in the video

For each miniclip, we store the extracted actions and their corresponding labels:

• 0 for visible
• 1 for not visible.

The visibile actions were manually cleaned by removing extra words like: usually, now, always, I, you, then etc. Example format in JSON:

{
  "4p1_3mini_5.mp4": [
    ["smelled it", 1],
    ["used this in my last pudding video", 1],
    ["make it smell nice", 0],
    ["loving them", 1],
    ["using them about a week", 1],
    ["using my favorite cleaner which", 0],
    ["work really really well", 1],
    ["wipe down my counters", 0],
    ["wiping down my barstools", 0],
    ["using the car shammies", 0]
  ]
}

Please cite the following paper if you find this dataset useful in your research:

@inproceedings{ignat2019identifying,
    title = "Identifying Visible Actions in Lifestyle Vlogs",
    author = "Ignat, Oana and Burdick, Laura and Deng, Jia and Mihalcea, Rada",
    booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = "7",
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
}

Some small parts of it are still under revision; please let me know here or by e-mail how I can help.

To download Stanford-postagger-full-2018-10-16 and all the required libraries. You need Python 3 (I have Python 3.6.7), it doesn't work with Python 2. Comment Tensorflow (for cpu) in requirement.txt if you use tensoflow-gpu instead.

sh setup.sh

Download glove_vectors.txt (pre-trained POS embeddings on Google N-gram corpus using POS info from 5-grams).
Download glove.6B.50d.txt embeddings. Put both of them in data.

There are 3 main modules: Youtube processing, AMT processing and Classification. The first 2 modules are still under revision. The third module can be used without the first 2 ones, as all the data is accessible from the data folder or Google Drive.

Needs an youtube downloader API key. Given channel ids - now 10, and playlist ids for each channel ( 2 playlists / channel), it downloads all the videos from each playlist. The code can be found in youtube_preprocessing.

python main_youtube.py

Does all the processing related to AMT (read data, spam removal, compute agreement). The code is in amt.

python main_amt.py

Everything related to classification models, embeddings and features can be found in classify.

The available models are: svm, lstm, elmo, multimodal (video features + elmo embeddings).

To call the methods, for example lstm:

python main_classify.py --do-classify lstm

The Extra data consists of: context and POS embeddings, and also concreteness scores for each word in the actions.

You can find the context information for each action in data/dict_context.json: each action is assigned the sentence it is extracted from. The sentences are extracted from the Youtube transcripts, using the Stanford Parser.

You can find both the POS and context embeddings in data/Embeddings. They consist of averaging the surrounding 5 left and right glove50d word embeddings. For future work, we want to use elmo embeddings.

The concreteness dataset from Brysbaert et al. can be find in data folder. Also, the data extracted from the file (just the unigrams and their concreteness scores) is in data/dict_all_concreteness.json.

The concreteness and POS of all the words in the actions is stored in data/dict_action_pos_concreteness.json.

To add these extra features to your model: for example run svm with context and pos embeddings:

python main_classify.py --do-classify svm --add-extra context pos

To run multimodal with concreteness:

python main_classify.py --do-classify multimodal --add-extra concreteness

The video features are Inception, C3D and their concatenation. These are found in data/Video/Features. By default, the multimodal model is run with the concatenation of features: inception + c3d.

To run multimodal with inception:

python main_classify.py --do-classify multimodal --type-feat inception

After running YOLOv3 object detector on all the miniclips, all the results are stored here. Copy them in data/Video/YOLO/miniclips_results.

For all of this data, there is code available to generate your own data also.

Look in main_classify.py parse_args method for the rest of the models and data combinations.