LSFD: Lightweight Single Stage Masked Face Detector with a CPU Real-time Speed
Youngsam Kim, Jong‐hyuk Roh, Soohyung Kim · 2021 International Conference on Information and Communication Technology Convergence (ICTC) · 2021
Although a large number of face detectors have been developed, one of the remaining challenges is to achieve CPU real-time speed with minimizing performance drop. To address this challenge, we propose a lightweight masked face detector named LSFD. The proposed model is inspired by FaceBoxes yet it is designed to be faster and more accurate. We proceed with various performance evaluations with three datasets: FDDB, WIDER FACE, and MAFA. The proposed model achieves superior accuracy in WIDER FACE validation set and competitive accuracy in FDDB. In addition, we first evaluate class-specific accuracy of masked face in MAFA test set. Our model runs at 46 FPS on CPU ([email protected]) for VGA-resolution and 37 FPS on Samsung Galaxy Note20 for 320×320 image.