Face detection research based on a tilt-angle dataset

Abstract

In the existing public face datasets, the horizontal frontal and left-right rotation poses are the majority, and the models trained by them can not meet the requirements of face detection in the overlooking situation. Aiming at this phenomenon, the Tilt-angle face dataset TFD is cited and further expanded, and the Tilt-angle face dataset TFD-B is manually collected. The RetinaFace algorithm is adopted to carry out multiple face detection experiments. Typical experiment A shows that compared with WiderFace, the average detection precision of TFD+TFD-B as training set is improved by 4.81% when looking down at 15°, 9.87% when looking down at 30°, 10.56% when looking down at 45°,12.63% when looking down at 60°, and 15.62% when looking down at 75°, which indicates that TFD+TFD-B can effectively improve the precision of face detection in the overlooking situation. At the same time, the experiments carried out further show that expanding the training dataset can improve the precision of face detection. TFD+TFD-B can be obtained at https://github.com/huang1204510135/DFD.