Shulou(Shulou.com)05/31 Report--

In this article, the editor introduces "what is CBNet" in detail, the content is detailed, the steps are clear, and the details are handled properly. I hope this "what is CBNet" article can help you solve your doubts.

Guide reading

So far, the object detection network that performs best on COCO data sets is CBNet, with an average accuracy of 53.3 on COCO test data sets.

Merging a more powerful backbone can improve the performance of the target detector. In order to do this, they propose a new strategy to assemble multiple identical backbones through composite connections between adjacent trunks. By doing so, they proposed a more powerful backbone called the composite backbone network.

As shown in the figure, an CBNet consists of multiple identical backbone networks and composite connections between adjacent backbones. From left to right, the output of each stage is in an auxiliary backbone, which can also be seen as a high-level feature. The output of each feature layer flows to the parallel level of the subsequent trunk through a combined connection as part of the input. By doing so, multiple high-and low-level features are fused together to produce richer feature representations.

This paper introduces two architectures: double backbone network (DB) and triple backbone network (TB). As you can guess from the naming, DB consists of two identical trunks and TB consists of three identical trunks. Performance differences are discussed later in this article.

In order to combine multiple outputs from the trunk, a combined connection block is introduced in this paper. The block consists of a 1x1 convolution and a batch normalization layer. These layers are added to reduce the number of channels and perform up-sampling operations.

The final trunk (on the far right of the figure) is called the leader trunk and is used for object detection. The output characteristics of the leader trunk are input to the RPN/ detection head, and the output of each auxiliary trunk is input to the adjacent trunk.

Combination style

There are four forms of backbone combinations:

The adjacent high-level combinations are the styles described in the previous section. Each output feature from the secondary backbone is input to the adjacent backbone using a composite connection block.

The same-layer combination is another simple composition style that provides the output of the adjacent lower-level phases of the previous trunk to the latter trunk. As shown in the figure, this style does not use composite connection blocks. Features from the lower backbone are added directly to the adjacent backbone.

The adjacent lower-level combinations are very similar to AHLC. The only difference is that the features from the bottom of the previous backbone are passed to the subsequent backbone.

Dense high-level combinations are inspired by DenseNet, with each layer connected to all subsequent layers, establishing a dense connection at one stage.

The above table shows the comparison of different combination styles. We can observe that the AHLC style is superior to other composite styles. The reason behind this is well explained in the paper. The author believes that adding the low-level features of the former trunk directly to the high-level features of the subsequent trunk will damage the semantic information of the latter trunk. On the other hand, adding the deep features of the former trunk on the basis of the shallow features of the subsequent trunk can enhance the semantic information of the latter trunk.

Result

The above table shows the test results of the MS-COCO test data set. Columns 5-7 are the results of object detection, and columns 8-10 are the results of instance segmentation. It clearly shows that the use of more backbone architecture improves the performance of the network.

After reading this, the article "what is CBNet" has been introduced. If you want to master the knowledge points of this article, you still need to practice and use it yourself. If you want to know more about related articles, 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.