Face and gait provide complementary information for person recognition, but their usefulness changes as a person approaches a camera: facial detail increases while the body progressively leaves the field of view. To support further research, we introduce Walk2Cam, a dataset of real-world approach videos containing 43 participants and 984 walks across five appearance conditions and three nominal walking speeds.We accompany this dataset with a reproducible benchmark comparing two face recognizers (GhostFaceNet and AdaFace), one gait recognizer (GaitBase), and each face recognizer fused with GaitBase. Evaluation uses complete walks and synchronized one-second observations, with identity-disjoint development and test sets. Using complete enrollment and probe walks at native resolution, both face recognizers correctly identify all test probes, while GaitBase reaches 97.46% balanced Rank-1 accuracy. A balanced medium-speed subset then tests the loss of spatial detail on one-second observations. At the strongest reduction, to an intermediate frame height of 270 pixels, AdaFace--GaitBase fusion achieves 75.08% Rank-1 identification accuracy, compared with 58.03% for the strongest individual modality. Automatic indicators of body framing and face detection success also identify subsets with lower recognition error, demonstrating a trade-off between accuracy and the proportion of observations retained. Together, Walk2Cam and its benchmark provide a testbed for studying when face and gait remain useful under changing acquisition conditions. The results show that fusion can substantially improve recognition from short, degraded observations, while highlighting the importance of observation duration and failure accounting in evaluating multimodal systems.

Assessing Visual Quality for Face-Gait Recognition: A Real-World Video Dataset and Benchmark / Boscolo, F., De Mola, G., Lamberti, F.. - ELETTRONICO. - (In corso di stampa). (2027 IEEE 17th Annual Computing and Communication Workshop and Conference (CCWC) Las Vegas (USA) 4-6 January 2027).

Assessing Visual Quality for Face-Gait Recognition: A Real-World Video Dataset and Benchmark

Federico Boscolo;Fabrizio Lamberti
In corso di stampa

Abstract

Face and gait provide complementary information for person recognition, but their usefulness changes as a person approaches a camera: facial detail increases while the body progressively leaves the field of view. To support further research, we introduce Walk2Cam, a dataset of real-world approach videos containing 43 participants and 984 walks across five appearance conditions and three nominal walking speeds.We accompany this dataset with a reproducible benchmark comparing two face recognizers (GhostFaceNet and AdaFace), one gait recognizer (GaitBase), and each face recognizer fused with GaitBase. Evaluation uses complete walks and synchronized one-second observations, with identity-disjoint development and test sets. Using complete enrollment and probe walks at native resolution, both face recognizers correctly identify all test probes, while GaitBase reaches 97.46% balanced Rank-1 accuracy. A balanced medium-speed subset then tests the loss of spatial detail on one-second observations. At the strongest reduction, to an intermediate frame height of 270 pixels, AdaFace--GaitBase fusion achieves 75.08% Rank-1 identification accuracy, compared with 58.03% for the strongest individual modality. Automatic indicators of body framing and face detection success also identify subsets with lower recognition error, demonstrating a trade-off between accuracy and the proportion of observations retained. Together, Walk2Cam and its benchmark provide a testbed for studying when face and gait remain useful under changing acquisition conditions. The results show that fusion can substantially improve recognition from short, degraded observations, while highlighting the importance of observation duration and failure accounting in evaluating multimodal systems.
In corso di stampa
File in questo prodotto:
Non ci sono file associati a questo prodotto.
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/3018199
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo