IEEE 2015 Conference on

Computer Vision and Pattern

Recognition

Vision for Intelligent Vehicles & Applications Face-Off Challenge:

Dataset Creation and Balancing Privacy

Sujitha Martin and Mohan M. Trivedi Computer Vision and Robotics Research Laboratory

University of California, San Diego (http://cvrr.ucsd.edu)

Why Faces?

LISA Face Detection and Head Pose Dataset

De-Identification Method

Sample Recognition Rate

Gaze-zone Estimation Accuracy

One-Eye

5%

65%

Two-Eyes

8%

71%

A two part user study using human participants showed face recognition to be below chance yet preserving gaze information well above chance with the following deidentification filter designs.

A common pool of naturalistic driving data is necessary to develop and compare algorithms in order to improve driving safety but it comes at the cost of privacy invasion.

There is a need for deidentifcation filters which will protect the identity of drivers while preserving sufficient details to infer driver behavior (e.g. eye gaze, head pose and hand activity).

Yield to Privacy! It’s the Law.

References

[1] Sujitha Martin, Ashish Tawari, Erik Murphy-Chutorian, Shinko Y. Cheng, Mohan Trivedi, "On the

Design and Evaluation of Robust Head Pose for Visual User Interfaces: Algorithms, Databases,

and Comparisons,” Automotive User Interfaces and Interactive Vehicular Applications (AUTO-UI)

Conference, Oct. 2012.

[2] Ashish Tawari, Sujitha Martin and Mohan M. Trivedi, "Continuous Head Movement Estimator

(CoHMET) for Driver Assistance: Issues, Algorithms and On-Road Evaluations,” IEEE

Transactions on Intelligent Transportation Systems, 2014.

[3] Sujitha Martin, Ashish Tawari and Mohan. M. Trivedi, "Towards Privacy Protecting Safety

Systems for Naturalistic Driving Videos,” IEEE Transactions on Intelligent Transportation

Systems, 2014.

• Driver distraction (e.g. talking, and eating) and inattention (drowsiness, fatigue, etc.) are some of the prominent causes of crashes in the US. One way to observe driver behavior is through facial analysis.

• The Face-Off challenge is one of a three-part challenge in the Vision for Intelligent Vehicles & Applications (VIVA) aimed at building robust and accurate vision algorithms for facial analysis. It includes an annotated dataset of images of vehicle occupants from naturalistic driving.

• Goals of the challenge: 1) Present the issues and challenges in vision from real-world driving scenarios [1] 2) Highlight deficiencies in current approaches and progress the development of future

algorithms [2]

Like Challenges? Come Participate in the

VIVA Challenge! http://cvrr.ucsd.edu/vivachallenge

Dataset: Images from 23 videos, 20 vehicles, subjects of varying age, ethnicity and gender Challenges: Illumination, occlusion, image quality, camera perspective, motion blur, deformation, etc. Tasks: Robustly and accurately do face detection and head pose estimation