A Novel Pyramid Network with Feature Fusion and Disentanglement for Object Detection

Figure 2

The overview of the proposed FFAD cooperated with single-stage detector. The features of input images are first extracted by the backbone network, and then MF³M fuse these features through multiple paths. Finally, TaConv produces a multilevel feature pyramid. There are two parallel feature maps used to predict specific categories and regress precise boxes, respectively, at each level.