Baidu image segmentation 7-day punch-in training camp learning summary

Baidu image segmentation 7-day punch-in training camp learning summary

The 7-day punch-in training camp for image segmentation of Baidu paddlepaddle that participated this time is coming to an end. This course is lectured by two big cows, Mr. Zhu and Mr. Wu. Mr. Zhu has lectured for 4 days, mainly theory and handwriting Code practice, Mr. Wu gave a lecture for 2 days, mainly explaining the cutting-edge content, summarizing the learning process and learning experience this time.

1. The learning content of this check-in camp

This course is mainly composed of the following chapters:

1. The preview content
   leads novices to quickly get started with Python, Notebook, Debug, PaddlePaddle and PaddleSeg through several sections of experiments;

2. The overview
   of image segmentation introduces the definition of image segmentation:
   including still image and image panoramic segmentation and video panoramic segmentation, as follows:

   
   This section introduces the basis for the learning of this course:
   
   From this section, Mr. Zhu led us to write code. Here we have built a basic network model that can also work.

3. Detailed explanation of FCN full convolutional network
   The full name of FCN is Fully Convolutional Networks, that is, there is no Full Connection, and all use convolution. In this way, image segmentation can be realized on the basis of image classification.
4. The U-Net/PSPNet model
   
   adopts the U-shaped structure of the encoder and the decoder to make the input and output sizes unchanged.
   PSP Net = Pyramid Scene Parsing Network
   is simply: multi-scale, or in the words of the teacher, "look left and right" (observe contextual semantics), increase the receptive field.
   
   Hollow convolution is used here.
   Dilated Conv = Atrous Conv= Hollow Conv
   
   . The picture here is made by the teacher in Excel. Remember to read a Japanese book explaining deep learning. The formulas, derivation process and graphics are also They are all made using Excel, which shows that Excel is a really powerful tool.

The teacher took everyone to hand-write UNet and PSPNet, and left homework, so that everyone can realize the practical application of the network based on the completion of the network, but I can only copy the teacher's code, but can not extend the application:(

5. The DeepLab series
   introduces several versions of the DeepLab series and when they appeared, and the main content of the explanation is as follows:
   
   Compare the structure of several versions: The
   
   teacher takes you here to hand-write the basic structure of DeepLab.

6. 
   The basic concepts of graph convolution network algorithm graph convolution are as follows:
   

7. 
   Introduction to image instance segmentation and panoramic segmentation introduces the concepts of semantic segmentation, instance segmentation, and panoramic segmentation:
   

2. Learning experience of this check-in camp

Through learning, I got an understanding of these concepts, and I followed the teacher to write the code. When I followed the teacher to write the code, I felt that the idea was very clear, but when I left the teacher’s video and wrote the code myself, my mind was still blank ( When I get older, I feel like my brain is like a piece of wood. I just can’t get the hang of it: (), there is no way to apply these contents to reality. Now my distress is that I can’t get in the door late, I can’t memorize the concepts, I can’t Writing programs by myself, it seems that I still have to continue to learn, and I must continue to participate in and learn other courses launched by Baidu aistudio.