In-The-Wild Face Angle Dataset

This dataset was created to be able to analyze the impact of aggregating embeddings with respect to a single setting and different settings. In particular, by having the dataset, three immediate questions could be answered:

  • How can we best aggregate different embeddings into a single embedding?
  • How do different types of enrollment-images affect face recognition accuracy?
  • What are potential challenges for self-enrollment?

Composition

The dataset consists of 11 people. The dataset contains different ethnicities, with 7 male and 4 female subjects with ages ranging from 22 to 85 years. Every person recorded a 3 to 6-second video starting with the left side of the face (profile), moving to the middle of the face (frontal) and ending with the right side of the face (profile again). We extracted the images from these videos. Additionally, we asked the participants to provide 10 additional images of themselves in completely different settings.

Collection process

We asked the participants to record the video themselves, in order to simulate a self-enrollment. There were no restrictions with respect to the device, although most participants used their phone. Participation was completely voluntary, and the subjects have not been financially compensated. The dataset was collected from December 2021 until May 2022.

Pre-processing

All images were converted to JPG. We extracted every frame from the received video.

Maintenance

The dataset is hosted by Johannes Kepler University Linz, Institute of Networks and Security (contact). We do not plan to update the dataset. However, in case a participant revokes their permission to use their facial image, the dataset will be updated.

All participants were informed about the data collection, as they signed a consent form and sent us their data. People can revoke their consent at any time.

Publications

If you want to use this dataset for your research, send the signed In-The-Wild Face Angle Dataset Release Agreement.