CoNAN: Conditional Neural Aggregation Network For Unconstrained Face Feature Fusion

Face recognition from image sets acquired under unregulated and uncontrolledsettings, such as at large distances, low resolutions, varying viewpoints,illumination, pose, and atmospheric conditions, is challenging. Face featureaggregation, which involves aggregating a set of N feature representationspresent in a template into a single global representation, plays a pivotal rolein such recognition systems. Existing works in traditional face featureaggregation either utilize metadata or high-dimensional intermediate featurerepresentations to estimate feature quality for aggregation. However,generating high-quality metadata or style information is not feasible forextremely low-resolution faces captured in long-range and high altitudesettings. To overcome these limitations, we propose a feature distributionconditioning approach called CoNAN for template aggregation. Specifically, ourmethod aims to learn a context vector conditioned over the distributioninformation of the incoming feature set, which is utilized to weigh thefeatures based on their estimated informativeness. The proposed method producesstate-of-the-art results on long-range unconstrained face recognition datasetssuch as BTS, and DroneSURF, validating the advantages of such an aggregationstrategy.