Source code and update details of new instance segmentation network panet (path aggregation network for instance segmentation)

PANet yes 18 An article in CVPR, The author is from Hong Kong Chinese , Peking University , Shangtang and Tencent Youtu ,PANET Can be regarded as Mask-RCNN+, Is in Mask-RCNN Several improvements based on .

PANet Source Baidu network disk ：https://pan.baidu.com/s/15H4XBAVXsz_lnwMXduXEBw
Extraction code ：frhs

PANet Post translation thesis

Path aggregation network （PANet）, It aims to improve the information flow in the proposed case-based segmentation framework . To be specific , Through bottom-up path enhancement, precise positioning signals are used in lower layers to enhance the whole feature hierarchy , Shorten the information path between the lower layer and the top-level function . We propose an adaptive feature pool , It links the feature grid to all feature levels , Make the useful information of each feature level spread directly to the following proposal sub networks . Supplementary branches that capture different views for each proposal are created to further improve mask prediction .

Improving direction ：
1、 In order to shorten the information path and enhance the feature pyramid by using the accurate positioning signal existing in the lower layer , Create a bottom-up path enhancement .
Picture description ：
（a）FPN The backbone .（b） Bottom up path enhancement .（c） Adaptive function pool .（d） Box Branch .（e） Fully connected fusion . For the sake of simplicity , We omit （a） and （b） Channel dimension of feature mapping in .

2、 To restore the disconnected information path between each proposal and all functional levels , Develop adaptive function pool .

3、 To capture the different views of each proposal , With a tiny full connection （fc） Layer to increase mask prediction , These layers are similar to Mask R-CNN Originally used FCN Prediction with complementary properties by integrating these two views , The diversity of information will increase , Produce a better quality mask effect .

made coco2017 First place in instance segmentation , Second place in target detection .