Extremely Tiny Face Detector for Platforms with Limited Resources.

Face detection is a fundamental step for face analysis tasks. In recent years, deep learning-based algorithms in face detection have grown rapidly. Most neural networks are computationally expensive and rely on graphics processing units, falling to be applied in practical applications. This paper explores the principles of designing tiny models and proposes an extremely tiny face detector based on the tiny-YOLOv3 framework, introducing new network structures such as Cross-Stage-Partial-connections (CSP), depthwise convolution, and Spatial Pyramid Pooling (SPP). The number of parameters is less than 10k, and the storage is less than 50Kb by using half-precision float point (FP16) for each parameter. Furthermore, each layer’s peak memory usage is under 0.07MB, leading the model to be accessible to various platforms. The experiments on a subset of the WIDER FACE dataset and Open Images Dataset V4 (OID) show that the proposed face detector can achieve comparable performance to the more massive face detectors heavier in model size and floating-point operations.