Evaluating Human Performance in Dynamic Perspective Invariant Face Recognition

Lim, Rosary; Ragab Sayed, Mona; Mandal, Bappaditya; Teck Ma, Keng; Li, Liyuan; Hwee Lim, Joo

Abstract

The aim of this study is to investigate and derive plausible consistent eye gaze scan path, and set of facial features learnt from unfamiliar faces in unconstrained dynamic motion (rigid and non-rigid motion) for subsequent recognition tasks using psychophysical experiments. Existing literature reported a shared observation that face recognition performance, in terms of accuracy for face verification and identification, are enhanced when subjects have been trained using faces in dynamic motion compared to when they only learn with static face images (Knight and Johnston, 1997; Lander and Chuang, 2005; Xiao et al., 2013). Although prior work suggested that dynamic motion provides additional information with an increase in number of frames to the identity of a face than static images (O’Toole et al., 2002; Schultz et al., 2013), the type of additional information (e.g. facial features, gaze scan path) underpinning that conclusion has not been discovered and explained. Our experiments aim to identify such features that are learnt by human subjects from dynamic motion to gain insight on the possible strategies engendering human’s superior performance across difficult verification conditions in unconstrained face recognition tasks, such as variations in illumination, viewing perspectives (poses), expressions, and age. Given that participants are generally unfamiliar with the face stimuli presented during the experiment, we are probing for generalized eye gaze scan paths and/or set of frequently fixated facial features for the respective verification condition. Such eye gaze scan path strategies and features will be evaluated later for potential translation into computational models in machines to emulate the competence of human recognition system in the hopes of improving current state-of-the-arts face recognition technology in the field of artificial intelligence.

Citation

Lim, R., Ragab Sayed, M., Mandal, B., Teck Ma, K., Li, L., & Hwee Lim, J. (2015, July). Evaluating Human Performance in Dynamic Perspective Invariant Face Recognition. Paper presented at 11th Asia-Pacific Conference on Vision (APCV), Singapore

• Presentation Conference Type: Conference Paper (unpublished)
• Conference Name: 11th Asia-Pacific Conference on Vision (APCV)
• Conference Location: Singapore
• Start Date: Jul 10, 2015
• End Date: Jul 12, 2015
• Deposit Date: Nov 17, 2023
• Publisher URL: https://easychair.org/smart-program/APCV2015/2015-07-10.html#talk:8221
• Additional Information: https://www.kybervision.com/iphone/apcv/index.php; https://www.kybervision.com/iphone/apcv2015/index.php