Deep Adaptive Feature Aggregation in Multi-task Convolutional Neural Networks

摘要: Convolutional Neural Network (CNN) based multi-task learning methods have been widely used in a variety of applications computer vision. Towards effective CNN architectures, recent studies automatically learn the optimal combinations task-specific features at single network layers. However, they generally construct an unchanged operation feature aggregation after training, regardless characteristics input features. In this paper, we propose novel Adaptive Feature Aggregation (AFA) layer for CNNs, which dynamic mechanism is designed to allow each task adaptively determine degree different tasks needed according dependencies. On both pixel-level and image-level tasks, demonstrate that our approach significantly outperforms previous state-of-the-art CNNs.