Papers Address: https://arxiv.org/abs/1612.03144
Although low-level features high-resolution pictures and more details, but the lack of semantic information; high-rise low-resolution map feature rich semantic information, but the effect on the detection of small targets and poor. FPN by Botteom-up transformation of the original, wherein FIG obtained contains both details and rich semantic information, can improve target detection.
A configuration diagram of FIG. 1 FPN
FPN basic mechanism shown in Figure 1, it can be divided into two parts:
1. Bottom-up pathway. Bottom-up path is the left image input process prior to the calculation CovNet, leftward in FIG. 1 wherein each layer feature pyramid FIG each level output of the last convolution layers, from the bottom, respectively referred to as a (C2, C3 , C4).
2. Top-down pathway and lateral connections. FIG FIG 1 wherein the right from the bottom three are referred to as (P2, P3, P4), wherein the topmost layer is characterized by P4 FIG C4 through convolution of a convolution kernel of size 1 * 1 is obtained (in order to make the right FIG characteristic dimensions are the same), then the bottom of the feature map are sampled by the upper layer and the characteristic diagram of the left results by corresponding features of FIG. 1 * 1 results obtained by adding the convolution. Finally, various features are on the right of FIG processed for target detection.
FPN mainly Patent Application Faster R-CNN RPN now in:
Instead of the original features of FIG. 1 with 5 wherein FIGS. Top-down of;
5 wherein FIGS. 3 * 3 convolution shared subsequent FC and two parameters;
anchor unset the RPN embodiment, corresponds to FIG. 1 wherein each scale, but may correspond to each scale 3 aspect ratio.