Determine the cause of poor cross-session generalization

Determine whether the inability of existing gait-focused recognition systems to generalize reliably across recording sessions is caused by limitations of the recognition systems or by the inability of the biometric data recorded by the sensors to support robust and stable cross-session identity inference.

Background

The MultiGait benchmark finds that existing recognition systems perform substantially worse when trained and tested across separate recording sessions than when evaluated within a single session. This limitation is especially pronounced for radar- and WiFi-based systems, although imaging-based systems and lidar also fail to achieve consistently reliable cross-session identification.

The paper does not resolve whether the observed failure reflects shortcomings in current recognition architectures or an intrinsic lack of stable, session-invariant biometric information in the sensor recordings. Resolving this distinction is important for designing improved recognition systems and for assessing the underlying privacy risks of multi-session sensor deployments.