piergiaj / avid

AViD Dataset: Anonymized Videos from Diverse Countries

AViD Dataset: Anonymized Videos from Diverse Countries

AViD is a large-scale video dataset published at NeurIPS 2020 (AViD NeurIPS details for video, poster and presentation information). It has 467k videos and 887 action classes. Importantly, AViD has several key attributes:

Static

The collected videos have a Creative-Commons License, allowing us to create and distribute a static dataset collected from various web sources (e.g., Flickr, Instagram, etc.). Unlike other YouTube-based datasets (e.g., Kinetics), the dataset is static and easily downloadable enabling reproducible research. We further release this dataset under a flexible MIT license, unlike more restrictive video datasets (e.g., Moments-in-Time and SomethingSomething). The dataset has similar size to the other standard video datasets.

Anonymized

All the faces in the videos have been blurred so that no person can be identified.

Diverse

The videos have been collected from a wide range of countries and sources. This is important as some actions, for example greeting, are performed differently in different cultures. Other actions, like news broadcasts, can have different text depending on the country. We find the model is unable to recognize videos from different countries without diverse training data (Tables 3, 4 and 5 in the paper).

Classes

The AViD dataset consists of 887 activity classes, capturing similiar actions to those in Kinetics, plus some additional actions such as talking, explosion, boating, etc. The classes follow a long-tailed distribution. More details on the classes and hierarchy of actions are described in the paper.

Baselines

Method	AViD Accuracy
2D ResNet-50	36.2
I3D	46.8
3D ResNet-50	48.2
Two-Stream 3D ResNet-50	50.1
RepFlow	50.5
(2+1)D ResNet-50	48.8
SlowFast-50 4x4	48.5
SlowFast-50 8x8	50.4
SlowFast-101 16x8	50.9

Dataset

The annotations are provided in this repository, in the dataset directory. This contains the full dataset with labels and weak tags as well as classification-only train and validation sets.

The videos can be downloaded. https://vision.cs.stonybrook.edu/~mryoo/avid/avid.tar.gz MD5 sum is available at https://vision.cs.stonybrook.edu/~mryoo/avid/avid.tar.gz.md5

Paper

AJ Piergiovanni and Michael S. Ryoo "AViD Dataset: Anonymized Videos from Diverse Countries" in NeurIPS 2020

arXiv

@inproceedings{aviddataset,
      title={AViD Dataset: Anonymized Videos from Diverse Countries},
      booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
      author={AJ Piergiovanni and Michael S. Ryoo},
      year={2020}
}

avid's Issues

Download links do not work

When I try to download the dataset from Drive, I see:

Google Drive seems like a bad place to store files for people to download. Some alternatives that I've seen:

• Dropbox
• GCS (Google Cloud Storage)
• AWS (Amazon Web Storage)
• bitTorrent (crowdsource the dataset distribution!)

Errors in avid_val.json

The action type of video fpzvtgor.mp4 is marked as "putting on mascara" which only appears once, I think it should be "putting on eyeliner".

Moreover, there are only 880 types contained in the validation set, the results will be more convincing if completed.

None of the Dataset Links work

I tried all the provided links, none of them work. The ones hosted on Google Drive have a "Quota exceeded" error, and the file that is split into three is corrupted (as reported before, not sure why it is still here). The last link to the small dataset (https://aviddataset.blob.core.windows.net/avid/avid_small.tar.gz) has the following error:
AccountIsDisabledThe specified account is disabled. RequestId:dd39b88e-801e-009c-56c5-7a4d98000000 Time:2020-08-25T09:52:50.3451882Z

Can you please provide a valid download link?

could you share the code that you used for 3D ResNet50?

Hi! I'd like to use the model that you used here, but couldn't find any link to the 3D ResNet 50. Have you uploaded the code somewhere?

Thank you! :)

anyone can download the dataset?

Dear all:
I tryied my ways to download the dataset, but none of them works. I wonder why the author published the paper but can not give a easy way to download. Anyone has the valid download link? Thank you very much.

Access to the code for "4D-Net for Learned Multi-Modal Alignment" in ICCV2021

Hi, Mr. Piergiovanni!

I read your manuscript entitled "4D-Net for Learned Multi-Modal Alignment" published in ICCV2021, and your work is excellent and inspiring!
To develop further research based on your valuable work, may I have the access to the code?

Thanks for your research! Looking forward to your reply.