Minimalist notes detnet:a backbone network for Object detection

Minimalist notes detnet:a backbone network for Object detection

The core of this paper presents a backbone network:detnet specially used for detection task. At present, the main method of Detection network is based on classification network plus FPN and RPN structure. Most of the classification networks increase pixel sensitivity by reducing the spatial size of the deep layer, and the smaller resolution has a certain degree of influence on the accuracy of the positioning of large objects. In this paper, the detnet of the rate=2 is used to replace the 3x3 convolution of the original bottleneck in the deep layer of the network, and to increase the accuracy of the detection, it increases the pixel sensitivity field without reducing spatial size and increasing the computation amount. In order to reduce the amount of calculation, the channel number in the deep part stayed at 256 without increasing.

The main comparison of the laboratory RESNET50-FPN structure, so that the corresponding DetNet59 structure, approximate structure to such as the above figure. Detnet is a transformation of the skeleton network, so it can be combined with the FPN structure. Because the DetNet59 calculation (flops-4.8g) is still more than the RESNET50 calculation (flops-3.8g), some people may say that the increase in performance is due to a larger amount of computation, so the article added ResNet101 (flops-7.6g) Contrast experiment, shown in the Det task A higher test result is obtained with less computational volume on the detnet.

But then the article is lazy, the rest of the analysis is only the contrast with RESNET50. In contrast to average precision (AP) and average recall (AR), the three-scale objects DetNet59 are more than ResNet50, and to some extent, it is very useful to improve the deep resolution. Of course, if this part is a ResNet101 contrast is more persuasive.

The dilated Res-block used in the article replaces the identity map with the 1x1 convolution in the cut section. A comparative experiment was also made to show that this is really useful.

Because Detnet is only a skeleton network, and can be extended to other types of networks, the article finally put DetNet59 into the MASKRCNN framework, display performance has improved.

The disadvantage of the article is that the DRN structure uses the same rate and does not cover all pixels, and it may be better to use some of the existing methods, such as rate=1,2,5 interleaving, to overwrite feature map with empty convolution.