Skip to content
[Preprint] "Privacy-Preserving Deep Visual Recognition: An Adversarial Learning Framework and A New Dataset"
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
imgs
LICENSE
README.md Update README.md Jul 24, 2019

README.md

PA-HMDB51

This is the repo for PA-HMDB51 (privacy attribute HMDB51) dataset published in our paper http://arxiv.org/abs/1906.05675.

This dataset is collected and maintained by the VITA group at the CSE department of Texas A&M University.

Overview

PA-HMDB51 is the very first human action video dataset with both privacy attributes and action labels provided. The dataset contains 592 videos selected from HMDB51 [1], each provided with frame-level annotation of five privacy attributes. We evaluated the visual privacy algorithms proposed in [3] on PA-HMDB51.

Privacy attributes

We carefully selected five privacy attributes, which are originally from the 68 privacy attributes defined in [2], to annotate. The definition of the five attributes can be found in the following table.

Attribute Possible Values Meaning
Skin Color 0 invisible Skin color of the actor is invisible.
1 white Skin color of the actor is white.
2 brown/yellow Skin color of the actor is brown/yellow.
3 black Skin color of the actor is black.
Face 0 No face Less than 10% of the actor’s face is visible.
1 Partial face Less than 70% but more than 10% of the actor’s face is visible.
2 Whole face More than 70% of the actor’s face is visible.
Gender 0 Cannot tell Cannot tell the person’s gender.
1 Male It’s an actor.
2 Female It’s an actress.
Nudity 0 The actor/actress is wearing long sleeves and pants.
1 The actor/actress is wearing short sleeves or shorts/short skirts.
2 The actor/actress is of semi-nudity.
Relationship 0 Cannot tell Relationships (such as friends, couples, etc.) between the actors/actress cannot be told from the video.
1 Can tell Relationships between the actors/actress can be told from the video.

Examples

Frame Action Privacy Attributes
brush hair skin color: white
face: no
gender: female
nudity: level 2
relationship: no
pullup skin color: white
face: whole
gender: male
nudity: level 1
relationship: no

Download link

Google drive

Label format

The attributes usually don't change that much across a video, so we only need to label the starting and ending frame index of each attribute. For example, if a video has 100 frames, and we can see a complete human face in the first 50 frames while a partial face in the next 50 frames, we would label [face: complete, s: 0, e: 49], [face: partial, s: 50, e: 99], where 's' is for 'starting' frame and 'e' is for 'ending' frame. Note that each attribute is labeled separately. For instance, if the actor's skin color is visible in all 100 frames in the same video (assume the actor is white), we will label [skin color: white, s: 0, e: 99]. The privacy attributes for all 'brush hair' videos are in brush_hair.json, similar with all other actions.

Citation

If you use this dataset, please cite the following

@article{wang2019privacy,
    title={Privacy-Preserving Deep Visual Recognition: An Adversarial Learning Framework and A New Dataset},
    author={Wang, Haotao and Wu, Zhenyu and Wang, Zhangyang and Wang, Zhaowen and Jin, Hailin},
    journal={arXiv preprint arXiv:1906.05675},
    year={2019}
}

Acknowledgements

We sincerely thank Scott Hoang, James Ault, Prateek Shroff, Zhenyu Wu and Haotao Wang for labeling the dataset.

Reference

[1] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre, “Hmdb: a large video database for human motion recognition,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2011, pp. 2556–2563.
[2] T. Orekondy, B. Schiele, and M. Fritz, “Towards a visual privacyadvisor: Understanding and predicting privacy risks in images,” in Proceedings of the IEEE International Conference on Computer Vision(ICCV), 2017, pp. 3686–3695.
[3] Z. Wu, Z. Wang, Z. Wang, and H. Jin, “Towards privacy-preservingvisual recognition via adversarial training: A pilot study,” in Proceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 606–624.

You can’t perform that action at this time.