| AI base of science and technology (rgznai100)
Participate in | Zhou Xiang, ShangYan
What exactly is the significance of this paper?
A very difficult problem in target detection and location is how to select the object containing the target from tens of thousands of candidate Windows. Recall of the model can be guaranteed only when there are enough candidate Windows.
At present, there are two main target detection frameworks:
One is One-stage, such as YOLO, SSD, etc. These methods are fast, but the recognition accuracy is not as high as two-stage. One of the important reasons is that it is difficult to use a classifier to not only suppress negative samples, but also classify targets well.
Another target detection framework is two-stage, represented by Faster RCNN. Such methods have high identification accuracy and positioning accuracy, but have problems of low computational efficiency and large resource occupation.
Focal Loss solved this problem from the perspective of optimization function, and the experimental results were very solid, excellent work.
In other words,One-stage detectors are faster and simpler, but not very accurate. Two-stage detectors are accurate, but expensive.
Results demonstrated that when using Focal Loss training, RetinaNet not only matched the detection speed of one-stage detectors, but also outperformed all current state-of-the-art two-stage detectors in accuracy.
Abstract
1 introduction
2 Focal Loss
3 RetinaNet detector
4 training
Table 1: RetinaNet and Focal Loss Stripping test (All-out Experiment)
5 test
Table 2: Target detection single model results (bounding box AP) VS COCO test-dev state-of-the-art method
6 the conclusion
Table 3: FL and FL* VS CE (cross entropy) results
The paper addresses
https://arxiv.org/abs/1708.02002