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

This article is about how to use marginal monitoring information to accelerate Mask R-CNN instance segmentation training. The editor thinks it is very practical, so I share it with you. I hope you can get something after reading this article. Let's take a look at it.

Mask R-CNN is a classic model of case segmentation. The author achieves faster network convergence by adding a new task to the Mask R-CNN framework.

This paper adds a new prediction task for MaskR-CNN, called Edge Agreement Head (perhaps translated as "Edge Protocol predictor"? Inspired by the way human instances are annotated When people annotate the instance at the pixel level, they will only pay attention to the edge of the instance, while the internal part of the instance only needs to simply copy the edge annotation information. So the mask edges of the instance are very useful, and they represent the instance very well. The function of Edge Agreement Head is to encourage the edge of mask predicted by deep network training to be similar to that of groundtruth.

Algorithm thought

By observing the prediction images output in the early stage of Mask R-CNN training, the author finds that most of the time the edges are not on the point, and it is obvious that the neural network is taking a detour.

Take a look at the following example:

This is some of the predicted Mask in the early days of Mask R-CNN 's in-depth network training, and found that it did not find the edge first, as humans did, and even went too far (you can predict it not very accurately, but at least show that you are working in this direction! ).

In order to avoid the detours of the neural network, the author takes the edge information of the example as a supervisory guide, that is, the groundtruth carries out edge filtering to let the neural network predict the edge of the example at the same time. Pointed the way.

Mask R-CNN 's multitasking loss function:

To do this, add a new branch, predict the edge and compare it with the edge of groundtruth, see the following figure

The author only does the above operation within the 28-28 size area of each instance (so the amount of calculation is limited). The edge of prediction and groundtruth is calculated by adding simple 3-edge detection, because edge detection is often used with image smoothing, so the image on the right adds a smoothing step.

In the figure above, Lp represents the way to calculate the difference between the two, as follows:

P represents the power parameter of the pixel difference.

The author tries common Sobel filtering and Laplacian filtering to detect edges.

The author adds a loss function through Edge Agreement Head, the complexity of the model increases slightly, without adding any additional model variables that need training, the computational cost of training increases very little, but the amount of calculation does not increase when network inference.

Experimental results

The author makes experiments on the MS COCO 2017 data set to compare the COCO AP metrics accuracy of the benchmark model and the proposed model when the benchmark model reaches 160k steps.

Table 1 shows that when the training reaches 160k steps, the model training using Edge Agreement Head achieves higher accuracy, especially the model using Soble edge operator.

Table 2 shows that acceleration without image smoothing is more obvious and achieves higher accuracy.

The comparison of the forecast results is shown in the figure:

Table 4 shows that higher accuracy can be achieved with longer training time and using Edge Agreement Head.

The above is how to use edge monitoring information to accelerate Mask R-CNN instance segmentation training. The editor believes that there are some knowledge points that we may see or use in our daily work. I hope you can learn more from this article. For more details, please 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.