FACE DETECTION USING REFINED-RETINAFACE MODEL

RetinaFace is a multi-task and single-stage face detection model that detects faces and landmarks. However, it has limitations in detecting non-face content in output bounding boxes and mislocalizes facial landmarks for profile faces. To address these issues, Refined-RetinaFace (R-RetinaFace) is proposed. R-RetinaFace adds a post-optimization module that resizes bounding boxes and ensures all landmarks are within them. R-RetinaFace outperforms RetinaFace on SDUMLA-HMT and CASIA-3D-FaceV1 databases. On SDUMLA-HMT, R-RetinaFace achieves an ideal detection rate of 98.02%, a moderate detection rate of 1.32%, and a poor detection rate of 0.66%. On CASIA-3D-FaceV1, R-RetinaFace achieves ideal detection rates of 92.2%, moderate detection rates of 7%, and poor detection rates of 0.8%. In contrast, RetinaFace did not achieve ideal detection on both databases. It achieved only moderate and poor detection rates. On SDUMLA-HMT, RetinaFace achieves a moderate detection rate of 96.32% and a poor detection rate of 3.68%. On CASIA-3D-FaceV1, RetinaFace achieves a moderate detection rate of 83.9% and a poor detection rate of 16.1%. These results put R-RetinaFace a state-of-the-art method for face detection