MTCNN with Weighted Loss Penalty and Adaptive Threshold Learning for Facial Attribute Prediction

Facial attribute prediction is a meaningful but challenging task in computer vision, due to the demand of recognizing many attributes of a face image at a time. In this paper, we propose a novel multi-task convolutional neural network(MTCNN) with a weighted loss penalty scheme and an adaptive threshold learning algorithm to predict facial attributes by groups. In particular, the way of grouping leverages the spatial relationships among attributes and the design of MTCNN greatly decreases model complexities. The core idea of the weighted loss penalty is to make the prediction probability of an attribute appearing on a face close to ground truth. During validation stage, the adaptive threshold algorithm learns the threshold for every attribute prediction adaptively. Experiments on two benchmark datasets, show that our method achieves superior performance to the prior state-of-the-art methods under two different evaluation protocols.