EfficientFaceV2S: A lightweight model and a benchmarking approach for drone-captured face recognition

Abstract

Face recognition in aerial imagery encounters distinctive challenges, including low resolution and varying pitch angles. The influence of image quality, particularly resolution, on the performance of existing face recognition systems is well established. However, limited exploration exists regarding the performance of FR models on drone-captured images, primarily due to the scarcity of suitable datasets. To address this gap, our study investigates the efficacy of face recognition models when applied to drone-captured facial images. We utilize state-of-the-art lightweight and large models, evaluating them across three drone-captured benchmarks and one dataset focused on low resolution. To ensure a comprehensive evaluation, we additionally adopt seven widely recognized benchmarks, which are artificially downsampled and rotated to simulate the impact of distance and altitude on the view from a drone to a target. Our results highlight a substantial decrease in accuracy across all FR models in these challenging scenarios. In response to this challenge, we introduce a model, EfficientFaceV2S. The proposed EfficientFaceV2S model demonstrates consistent performance across all benchmarks while imposing minimal computational demands. This makes it particularly suitable for real-time and resource-constrained applications. The significance of our work lies in the development of EfficientFaceV2S, which effectively addresses the unique challenges posed by drone-captured images, offering significant improvements in accuracy and efficiency over existing models.