PaSC Logo

Point-and-Shoot Face Recognition Challenge
BTAS 2016 Video Person Recognition Evaluation

Overview of Competition
Person recognition in video remains a relatively unexplored problem with many open challenges and questions. This is particularly true for videos of people carrying out activities, as opposed to people speaking to a camera. This evaluation will include videos from the Point-and-Shoot Face Recognition Challenge (PaSC) and the Video Database of Moving Faces and People (VDMFP) collected at the University of Texas at Dallas. The PaSC data set was collected at the University of Notre Dame and includes multiple videos of 265 people carrying out relatively ordinary actions such as picking up an object or throwing a ball. This data set was gathered to spur the development of algorithms that find the relevant information and recognize the people in videos where people are engaged in some activity not directly associated with the camera. There is a strong presumption that much of the current work on this task will involve face recognition, but it is not a requirement.

The PaSC videos were used in the IJCB 2014 Handheld Video Face and Person Recognition Competition and the FG 2015 Video Person Recognition Evaluation. The VDMFP videos were used in the Video Portion of the Multiple Biometrics Grand Challenge (MBGC). Importantly, human performance benchmarks exist for both the PaSC video challenge and the VDMFP.

There are two innovations over previous PaSC-based competitions. The first measures algorithm performance on two video datasets collected at different institution. By incorporating two qualitatively different dataset, the competition will measure the ability of algorithms to generalize across datasets. The second will compare human and algorithm performance on videos on two datasets.

Update on UTD problem: The UTD problem consists of one similarity matrix and corresponding mask matrix. The source matrices can be found on the support page. To compute the similarity matrix, compare all the videos in the target set against all videos in the query set. To compute performance use the UTD mask matrix.

Call for Participation [PDF]

Important Dates:
Evaluation announcement: February 1, 2016
First round similarity matrices delivered and summary of approach given to Notre Dame: April 15, 2016
Final similarity matrices delivered to Notre Dame and option to supply modified approach description: May 4, 2016
Updated report delivered to BTAS 2016: May 10, 2016
Final notification of BTAS 2016 decision on report: June 10, 2016

Patrick J. Flynn (Notre Dame)
Walter J. Scheirer (Notre Dame)
P. Jonathon Phillips (NIST)