Deecamp32

============

This is the Projet to generate the closed boundary of a image we have tried many difference methods to slove it we did not reach the point we except, but get some results.

HED:

code: https://github.com/s9xie/hed
paper: http://openaccess.thecvf.com/content_iccv_2015/html/Xie_Holistically-Nested_Edge_Detection_ICCV_2015_paper.html

we use the pretrained model directly to our project, we get excited at first, Because it gives us a lot of confidence, he has a much better effect than the traditional boundary detect operator(such as Canny, Sobel), and it is also firm that we use a learning-based approach to solve the problem.Results are as follows:

. . .

Although it has achieved relatively good results, our project requires that the external wireframe be closed, or find a way to detect the unclosed point, which brings us some trouble, outside the wireframe based on the hed method. The border is very thick, we use the skeleton algorithm to thin the line, and then use the power-off detection algorithm, which can solve the problem better.

In order to prevent the outermost wireframe from closing, we also use an energy-based algorithm (snake, level set) to enclose the outer frame at the outermost layer, which also has a good effect.
code:https://github.com/USTCzzl/Deecamp27/blob/master/image/levelset/level_set.cpp

RCF

code:https://github.com/yun-liu/rcf
paper: http://openaccess.thecvf.com/content_cvpr_2017/html/Liu_Richer_Convolutional_Features_CVPR_2017_paper.html

BDCN

code:https://github.com/pkuCactus/BDCN
paper:https://arxiv.org/pdf/1902.10903.pdf

Both of the methods we have tried, we retrained the model, but the results is not satisfied, there are some enhance but not enough, the three method almost some, only some difference in network structure,i will put my code in github.now let me show some difference way to thinking the problem.

Pip2Pix & Cycle Gan

this is a method from Gan, but someone has used it to do the edge search problem the paper is here: https://arxiv.org/abs/1901.00542
code: https://github.com/mtli/PhotoSketch it indeed get progress in BSDS500, but it lack of generalization in our datasets, the result is very abstract.but style transfer can work in this problem, our group member use cycle gan train the dataset about 4 days get wonderful results, i will show some picture in the future..

code

if you want to test in your own image, you can download the test document(only HED && RCF),there are three different methods.