Multi-Task Network and Optimization for Face Detection and Attribute Analysis

More and more application scenarios require algorithms to be able to detect human faces while predicting facial attributes such as gender and age. However, the existing face detection and facial attribute analysis are generally been solved as separate problems. At the meantime, how different tasks influence each other and how to balance and optimize multiple tasks still need to be further studied. Therefore, we design a novel multi-task network to jointly detect faces and predict facial attributes. Furthermore, we propose an optimization method based on noise estimation to adaptively tune the multi-task loss weights. Experimental results on CelebA dataset show that our method achieves great performance in both accuracy and speed.