Dynamic Feature Fusion for Visual Object Detection and Segmentation

Feature fusion is a key process of integrating multiple features in deep neural networks (DNN). The mainstream method in the literature is based on the Feature Pyramid Network (FPN), where the learned parameters about feature fusion is fixed after the training process. That is, how the multiple features will be fused is independent from the embedded characteristics of the input data, making the feature fusion process less flexible especially for the object categories less seen in training data. Therefore, this paper proposes a novel feature fusion mechanism, called dynamic feature fusion. With this mechanism, a model can automatically learn and select the appropriate way of feature fusion to provide prediction heads with more effective and flexible input features depending on the characteristics of input data.