A thesis
Feature Selective Anchor-Free Module for Single-Shot Object Detection
https://arxiv.org/abs/1903.00621
https://www.cnblogs.com/fourmi/p/10602936.html
Second, the paper notes
1. Background
a), the author proposes a network structure FSAF blocks (feature selective anchor-free). This building block can be added to the network characterized in the pyramid to improve the detection accuracy. FSAF block network structure is mainly based on two issues addressed anchor object detection heuristic guide feature selection based sampling anchor 2. overlap. FSAF main role is to help our goal to find the most suitable to detect their characteristic scale.
B), previous work are artificial anchor box design size, and design rules artificial ground truth according to the size of the box, to find the corresponding size on the feature map to return FPN block, (such as a large target selection rear layer semantic information richer feature map, select small targets front layer details more feature map)
However, this method of artificial selection, not flexible enough, such as target 50 * 40 * 50 and 40 may choose a feature map, this may not be the best choice.
2, innovation
a), anchor-based method FPN Similar branches each feature map a anchor-free design, the training time for the feature map automatically selected according to the content of the target regression, estimated process may use a single anchor-free branch It can also be used with co-anchor-based branch.
b), the network structure
c), focal loss classification branch using regression branch using iou loss
d), in each instance by calculating each classification level feature map, and the return loss and selecting a smaller loss as a suitable feature map,
The intuition is that the selected feature is currently the best to model the instance
3, details
a), the training time would be a loss function loss anchor based anchor free and added together
b)、Ground-truth and Loss
This part of the reference https://blog.csdn.net/diligent_321/article/details/88384588
Classification Output
This place is classify each pixel, so the paper can also be viewed using anchor free branch inside similar semantic segmentation routine to do.
Box Regression Output
4, experimental
Experiments of this paper, the use of the enhanced test test, while the best results are added together anchor based branch of effect.