introduction
For small targets usually need to use multi-scale detection, FPN proposed by the authors is a quick and good effect multi-scale detection methods.
method
a, b, c is the previous method, in which a, c use the idea of multi-scale detection, but they all have significant drawbacks.
a method: every pictures are scaled, during testing, the biggest problem with this approach is too slow, because several times to spend time;
Method c: Method SSD paper is actually used, the sampling feature map layers, and between the different feature map scale to predict, the biggest drawback of this approach is high but the resolution of the underlying semantic feature map information weak classification are not allowed;
The new method proposed by the authors in the paper --FPN (characteristic pyramid network), with c speed as fast at the same time more accurate than c.
In fact, the principle is very simple:
We know that the underlying feature map semantic information but weak high resolution, low resolution, but the top feature map semantic information strong, so only the top-level semantic information will be passed to the bottom, you can let the classification more accurate.
Method of use, is the upper layer of the sample feature map to 2 times, and then added directly through the adjacent lower layer with feature map 1 * 1 conv is.
to sum up
FPN very large contribution to the paper, proposed to let the different layers of ideological feature map information fusion, widely later reference. For example, SSD upgrade DSSD, used deconvolution layer is the same principle, the effect is also very good.