Shulou(Shulou.com)06/01 Report--

Editor to share with you how to achieve the anchor free method in Segmentation, I believe most people do not know much about it, so share this article for your reference, I hope you can learn a lot after reading this article, let's go to know it!

3. Segmentation-based method FCOS, let's first take a picture of the overall framework of FCOS:

According to FCOS, it treats each location as a sample, as shown in the following picture, you can see that the leftmost orange dot is in the baseball player's box, and the gt of this point is actually the distance from this point to the four edges of box and the obj category of box, so the final predicted output is HxWxC and HxWx4. C and 4 represent the category to which the point to be predicted by each location of each feature graph belongs to and the boundary distance from the point to box, respectively. After the box is obtained in this way, the FCOS round anchor based method is the same as NMS and so on:

In fact, if you don't consider the Center-ness branch under Classification, will you feel very familiar? Post the network diagram of retinanet here:

It can be found that the biggest difference between the two is the final output channel. Retinanet outputs KA and 4A (A represents the number of anchors and K represents the number of categories), and predicts their category and relative offset for An anchors in each location location, while FCOS directly predicts the category of the lattice and the generation of box, which has no concept of box and is very close to the segmentation idea of semantic segmentation as a whole. There is a problem with this method, that is, in the box, the closer to the center, the better, but closer to the edge of the Box, although it should still be positive in theory, the prediction effect is not good because it often falls outside the obj. The solution of FCOS is to introduce a new branch centerness, whose gt is calculated as follows:

As you can see, if the left boundary distance and the right boundary distance between location and box are the same, the first term in the root sign should be 1. Similarly, when the distance between the upper and lower boundaries is the same, the second term in the root sign is 1. At this time, the GT value of 1 is exactly in the center position. If the location is very close to the edge, the gt will be very small. After this branch is trained, the inference phase will be multiplied by the predicted value of classification as the final scores score, thus suppressing the position close to the center point. In addition, FCOS also introduces the concept of multi-scale, if on a level of FPN, the maximum value in t/b/l/r is greater than a certain threshold, then the box is considered not suitable for the feature of the current level, so it is excluded. After Foveabox learned about FCOS, it was easier to get to know foveabox. First of all, the difference of foveabox lies in the multi-scale strategy and the encoding method. The multi-scale strategy of foveabox is to assign different sizes of box to different level feature map according to area, and there is overlap. The Pl of each leval of the P3~P7 of FPN has a cardinality Sl. When taking LF3, the S3 corresponding to P3 is 3232, and when taking LF4, the S4 corresponding to P4 is 6464, which is doubled all the time. The area range of the box that each level is responsible for is, where n ^ 2 is a variable parameter. You can see that the ranges predicted by different leval overlap, which can increase some robustness:

Taking into account the different sizes of box predicted by different level, the predicted box position coordinates are also encoded as follows (z represents coefficient, see paper for details):

Finally, that is the origin of the name of foveabox. For the suppression methods within box that are far from the central point, foveabox does not have the same branch as centerness, but uses another idea, that is, only the point within the box that is closer to the center is regarded as a positive sample (the red area with black dots below). If the point is inside the Box but close to the edge, it is often regarded as a gray area, which is neither a positive sample nor a negative sample. Gradient return is not considered (the white area inside the red box). The size of the rectangular box in which the positive sample is located and the rectangular box in the gray area are controlled by two different expansion coefficients.

FSAF doesn't have much to say about FSAF, just need to know the following three points: 1. The way FSAF predicts box in each location is also the distance between the predicted point and the four boundaries of box, similar to FCOS 2. FSAF suppresses the interference far from the central point inside the box similar to Foveabox, which also regards the rectangular area very close to the center as a positive sample. The multi-scale strategy of FSAF, which is relatively far away in box, does not take into account the gradient calculation. FSAF's multi-scale strategy is special. It does not manually assign level to predict a certain box (unlike FCOS and Fovea, which essentially use the method of manual allocation). Instead, it is calculated at the same time on each scale to see which level has the smallest loss of box, which is calculated on the feature of this level. As a result, automatic feature selection for different obj is realized.

The above is all the content of the article "how to implement the anchor free method in Segmentation". Thank you for reading! I believe we all have a certain understanding, hope to share the content to help you, if you want to learn more knowledge, welcome to follow the industry information channel!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.