【Background】：ECCV is one of the top conferences in computer vision,In this blog,I will introduce an paper from Sun Jian team, which is about a backbone network for object detection.What is worth mentioning is that this paper does not have any formula.
paper name: DetNet: A Backbone network for Object Detection
paper url: https://arxiv.org/abs/1804.06215
paper-information
1、【Abstract】
Object detection is a heavily researched topic in computer vision. It aims at finding “where” and “what” each object instance is when given an image.
In network structure, recent CNN based detectors are usually split into two parts. The one is backbone network,and the other is detection business part. you can see the picture below to understand this.
1.png
2、【Motivation】
the existing backbone networks usually have some problems,because they are desighed for classification task at first,there is no doubt that there are differences between different tasks(classification and detection).so,de signing a new backbone networks for detection is become very neccessary.
2.png
3、【The problems of exist backbone network】
2.png
4、【Comparisons of different backbones used in FPN】
3.png
5、【Contributions】
4.png
6、【DetNet design】
5.png
6.png
7、【More details】
the author adopt ResNet-50 as baseline, which is widely used as the backbone network in a lot of object detectors. To fairly compare with the ResNet-50, we keep stage 1,2,3,4 the same as original ResNet-50 for our DetNet.the more details you can see picture 7.png.
7.png
apply bottleneck with dilation as a basic network block to efficiently enlarge the receptive filed. Since dilated convolution is still time consuming,our stage 5 and stage 6 keep the same channels as stage 4 (256 input channels for bottleneck block). This is different from traditional backbone design,which will double channels in a later stage.
More information about dilation:https://blog.csdn.net/jzrita/article/details/72639969
8、【Experiment】
4.png
5.png
6.png
7.png
8.png
9.png
10.png
11.png
12.png
13.png
14.png
