A densely connected feature pyramid network for object detection

Feature pyramid plays an important role in object detection. Feature Pyramid Network(FPN)[1] looks perfect, but still has some short-comings: Rough coarsest nearest neighbor interpolation is used, so that the high-level feature information may not be effectively propagated. In the paper, we propose more efficient dense feature pyramid network structure, which call Dense-FPN. Our architecture essentially adds a series of dense skip pathways for FPN. Dense-FPN is essentially different from feature pyramid network (FPN)[1]: 1) has a dense skip connection on the skip pathways, which improves the gradient flow and enhances the semantic information; 2) uses a dense skip connection to enhance the transfer of features, the semantic information between the down-sampling and up-sampling maps is bridged. Extensive evaluation of multiple data sets shows that under different evaluation indicators, the performance of the model has achieved good performance.