Privacy-Preserving Deep Visual Recognition: An Adversarial Learning Framework and A New Dataset

We have recently introduced PA-HMDB51 (https://github.com/htwang14/PA-HMDB51), the very first human action video dataset with potential privacy leak attributes annotated. This dataset is collected and maintained by the VITA group at the CSE department of Texas A&M University.

The dataset contains 592 videos selected from the HMDB51 dataset [2]. For each video, we provide with frame-level annotation of five privacy attributes: skin color, gender, face, nudity, and relationship. The annotations are all provided in JSON format. Visualized examples can be found in the attachment.

The dataset aims to support and promote research on protecting visual privacy information in smart camera-based applications. A manuscript [1] introduces the dataset and related algorithms that we have developed for this topic.

We hope you will find this dataset useful,

Haotao Wang,
Texas A&M University