Purpose
Using multi-patches at present, there are two purposes, one is to maintain the original resolution, and the use of multiple patch to represent the originals, and the other is to use patch to deal with fine-grained fine-grained features.
multi-patch combination
Although also cut out the original image of the content, and object detection network is different, many patch more emphasis on the relationship between each other, so will patch combination again. The combination of the way is also followed.
1. The simplest way to use a simple combination of concatenate
2. Complex, if there are global and local characteristics, you can use the door structure of the way, using the global filter local characteristics.
3. There are also purposeful, such as looking at the quality of the picture or clear this information, comprehensive use of various statistical methods to get the characteristics of filtering, and then concatenate them together. There's also the sort of vector combination that's grouped together.
4. Also, to group patch, view the importance of grouping to the final feature performance to give different weights.
It can be seen that the combination of patch, according to the semantic information context and the information content of the block, to determine that there is no very cool very common practice. But one thing is for sure, if the end result needs to be rooted in each patch, the patch will need to be grouped together.