ACCURACY - COMPLEXITY TRADE - OFFS OF FACE DETECTION ARCHITECTURES AND THE ROLE OF 2D AND 3D BENCHMARK DATASETS IN THEIR EVALUATION

Abstract

This article examines the performance-efficiency trade-offs inherent in both conventional and neural-network-based face detection using empirical findings from the WIDER FACE benchmark. The study introduces marginal-cost and saturation metrics as analytical tools and systematically categorizes two-dimensional and three-dimensional benchmark datasets by their scale, variability characteristics, and subject density to facilitate the informed selection of appropriate detectors for evaluation.