This paper is a minor modification of FCOS, and the final performance reaches 50.4AP, which can be said to be quite strong. You can refer to the improvement and improvement methods in the engineering

Source: Xiao Fei’s algorithm engineering notes public account

FCOS: A Simple and Strong Anchor-freeObject Detector

  • Address:Arxiv.org/abs/2006.09…
  • Thesis code:Github.com/aim-uofa/Ad…

Introduction

The paper has been revised and published by the original FCOS team. The author calls it an extended version of FCOS rather than FCOSv2, but let’s call it FCOSv2 for the sake of discrimination. FCOSv2 has the same overall idea as FCOSv2, but much better performance. Compared to FCOS, performance on the base backbone RESnet-101-FPN was improved by 43.2AP from 41.5AP, and the highest version was up to 50.4AP. This paper mainly discusses some improvement and promotion methods of FCOSv2. For other specific implementations, please refer to the previous FCOS article.

FCOSv2

FCOS

FCOSv2 is basically consistent with FCOS in thought. It predicts the target category, size information and center-ness at each position of the feature map. The center-ness is used to represent the distance between the current position and the target Center point, and the final score of the target is obtained by combining the classification score and the center-ness score. Different from the previous Bbox regression method, the size information predicts the distance between the feature position and the four boundaries of the target.

Figure 2 shows the backbone network structure in FCOSv2. The backbone network still adopts FPN, and each layer feature uses the common head prediction category information, size information and centerness. For details, please refer to the previous article.

Change

Some modifications of FCOSv2 to FCOS are listed below. Since the paper does not compare with the original version, it is unclear how much revenue each part brings:

  1. The modification specified by the positive sample point, FCOS requires that the feature point is located inside the target, and the distance from the feature point to the target boundary meets the constraints of the FPN layer. FCOSv2 requires that the feature points are located in the central region of the target (Cx − RS, Cy − RS, CX +rs,cy+ RS)(C_x-Rs, C_y-rs, c_x +rs, c_y+rs)(Cx − RS, CY − RS, CX + Rs, Cy + RS), SSS is the stride of the current layer, and r=1.5r=1.5r= 1.5r=1.5 is the superparameter.
  2. The regression target of FCOS is directly the distance from the feature point to the target boundary. Since the Head is shared, a learnable scale factor is preset for each level during prediction. However, the stride is added to FCOSv2 to become more adaptable to the size of FPN, and the learnable scale factor is still used.

  1. In terms of the location of the center-ness prediction, the center-ness prediction of FCOS is put together with the classification prediction, while FCOSv2 is put together with the regression prediction.
  2. The regression loss function is modified. FCOS uses IoU loss for regression learning, while FCOSv2 uses GIoU loss for regression learning.
  3. For the final fraction calculation, FCOS is the product of classification fraction and center-ness, and FCOSv2 is the square root of the product of classification fraction and center-ness:

Improvement

In order to obtain better performance, in addition to replacing a stronger backbone network, the paper also extended FCOSv2 as follows:

  1. BiFPN is used instead of plain FPN, note that deep separation convolution is not used, this part can bring about 2AP improvement.
  2. Data enhancement in the test phase, the image was scaled to [400,1200][400,1200][400,1200] in sequence, each step size was 100, each size was inferred using the original image and the vertically flipped image, which could bring about 2.5ap improvement.
  3. The addition of deformable convolution, which replaces the third and fourth stage convolution of the backbone network, and the first four convolution of the two branches of Head, respectively, resulted in a boost of about 1.5AP.

Experiment

Compare with SOTA method.

Inference performance comparison.

Conclusion

This paper is a minor modification of FCOS, and the final performance reaches 50.4AP, which can be said to be quite strong. You can refer to the improvement and improvement methods in the engineering.

 



If this post helped you, please give it a like or watch it again

More content please pay attention to the wechat public account