Source of modules: In order to better adapt the model to targets of different scales, the FPN structure is often added to the network. It has actually been proved that the complementarity of high and low-level information will indeed improve the detection performance. However, the author of the Libra R-CNN network believes that FPN is problematic, and it does not effectively utilize feature information.
Origin of module: Traditional FPN and FPN variants such as PAFPN, BiFPN, etc. all use Top-Bottom and Bottom-Top methods, which pay more attention to adjacent resolutions, and the semantic information contained in non-adjacent layers is in the information. During the fusion process, it will be diluted once or even many times. This process is imbalanced. Therefore, the BFP module is to balance the fusion process, and it is an extension of FPN and its variants.
Function of the module: Use the feature map information of multiple levels to enhance the expressive ability of each level of feature map. Combining the respective advantages of FPN and non-local solves the imbalance problem very well.
Implementation: