Table of contents
1. Disadvantages of Anchor-based
2. Disadvantages of Anchor-free
1. Disadvantages of Anchor-based
The quality of the anchor box design determines the upper limit of the performance of the detection algorithm to a certain extent.
1) Anchor-based detection performance is very sensitive to the size, aspect ratio and number of anchor boxes , which means that parameters should be adjusted carefully. However, many hyperparameters are used in the design of anchor boxes, which brings problems such as difficulty in parameter adjustment and inconvenient use.
The anchor box design has three hyperparameters:
- Anchor box association layer : which feature layers are selected as the detection layer for the associated anchor box
- The size of the anchor box: the scale size of the associated anchor box on each detection layer
- Anchor Box Scale: The aspect ratio of the associated anchor boxes on each detection layer
After the anchor frame is designed, the anchor frame needs to be matched during the training process to divide positive and negative samples. Anchor box matching has two hyperparameters :
- Select the IoU threshold of the positive sample: the anchor box greater than or equal to the IoU threshold is a positive sample
- Select the IoU threshold of the negative sample: the anchor box less than or equal to the IoU threshold is a negative sample
2) Even after careful parameter adjustment, due to the fixed size of the anchor box, its universality is greatly reduced . For different tasks and targets with large deformations ( different types , scenes , scales , and proportions of objects to be detected ), it may be necessary to artificially design and adjust the hyperparameters of the anchor frame based on experience. It is difficult for inexperienced novices to achieve relatively Good application effect. And when applying the anchor box-based detection algorithm, the most time-consuming part is the adjustment of the anchor box hyperparameters.
3) In order to obtain the highest possible recall rate for anchor-based algorithms, many anchor boxes are needed to prevent missed detection. But because most of the anchor boxes are negative, this will lead to a serious category imbalance .
4) Since the tagging strategy of the anchor box is to compare the IOU size, this leads to more complex calculations for such methods .
2. Disadvantages of Anchor-free
Anchor-based has disadvantages, and Anchor-free of course also has shortcomings. Most of these shortcomings are the advantages of Anchor-based.
1) The recall rate of the box is insufficient. Since many Anchor-free are only predicted at positions close to the center of the gt box, the quality of the regression box is average. Such as YOLOV1.
2) The detection effect for overlapping targets is average. Such as the DenseBox series.
Reference link:
How to evaluate the latest anchor-free target detection model FCOS? bzdww
[Target detection] Object detection without anchor frame - key point method CornerNet bzdww