Accenture-MM1: A Multimodal Person Recognition Dataset

Kyle O'Brien, Michelle Rybak, Jiong Huang, Adam Stevens, Madeline Fredriksz, Michael Chaberski, Danielle Russell, Lindsey Castin, Michelle Jou, Nishant Gurrapadi, Marc Bosch; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2024, pp. 112-122

In this paper we present a new dataset to fuel multimodal research in uncooperative and surveillance scenarios. Accenture Multimodality 1 (ACC-MM1) is a large-scale multimodal biometric recognition dataset composed of imagery and video. The dataset includes challenges such as long ranges, high pitch angles, varied atmospheric conditions, and mixed image quality levels. Ultimately, a dataset containing 227 unique subjects, 303 hours of video, and 12,344 still images was captured in indoor and outdoor conditions. In addition to traditional modalities (face, gait, etc.), data for a novel biometric modality, activity gait, was collected. Covariates included appearance changes, walking with weighted loads, and body distortions. Furthermore, to enable standardized performance testing of ACC-MM1, an evaluation protocol was created. Baseline performance of popular and novel recognition algorithms is reported to encourage research in the challenging conditions present in ACC-MM1.