Minimalist notes Path Aggregation network for Instance segmentation

Minimalist notes Path Aggregation Network for Instance segmentation

Paper Address https://arxiv.org/abs/1803.01534

The core of this article is to propose a network panet for instance segmentation, as shown in Fig. Three improvement:
bottom-up Path Augmentation

After FPN, add the convolution operation from bottom to top. The article explains that before the FPN skeleton network depth is very deep (to a 100+ layer), in order to facilitate the bottom of the information transmission, can be used at the decision-making level, so after the FPN added bottom-up path, so that the bottom of the information to the decision-making level of the shortest path from the Back-bone to go , from FPN cut to walk bottom-up path add up to less than 10 layers, can be more effectively used. Bottom-up path structure as shown in figure, N2 N 2 n_2 is P2 P 2 p_2, after which Ni N I n_i passes through a 3x3 convolution (stride=2), plus pi+1 P i + 1 p_{i+1}, and then a 3x3 convolution gets ni+1 n i + 1 n_{i+1}, so loop. Adaptive Feature Pooling

The author has made a statistic that the decision information of a certain scale is not entirely from the corresponding level of feature map, the vast majority of information from other levels, so all levels of feature map need to use. In this paper, the author proposal The roialigned of the feature map at different levels to make the max operation Pixel by point, which is the feature map fuse of different scales.
fully-connected Fusion

The author thinks that when the segmentation, the FCN way mainly relies on the local feeling field, and Fully-connect layer for many local information combination can have the better expression ability (the parameter is more), can combine each local information to form the proposal whole. So the author opens a new branch with a full join operation and maps to the same spatial size and FCN results added. The authors say that only one full connection, not multiple times, is to prevent the full connection from losing the spatial information of the feature. In the two feature map Fusion, the author experimented with "Max", "Sum", "product", and found that "sum" was good.

Panet Instance Segmentation 1st and Object Detection 2nd in Coco 2017 challenges, compared to the champion effect of previous years, the following table