Segmentation of structured light images using segmentation network

1. U-Net

github Connect

We use the pre training model provided by the web page unet_carvana Carry out the alignment carvana And speckle structured light images （ Including simulation and real images ） Carry out class II segmentation （ prospects - Background segmentation ）. The result is ：

• Yes carvana The effect of image segmentation is good , This is because the model uses carvana Data sets for training .
• The segmentation effect of speckle structured light is very poor , Especially for simulation images , It is completely impossible to split .

Our segmentation results are as follows ：

carvana Images

Structured light image

For simulation images , Test the same image projected by three different structured lights . For real images , For general 、 indoor 、 Test the structured light images in three outdoor environments .

• real-general
  
• real-indoor1
  
• real-indoor2
  
• real-indoor3
  
• real-outdoor
  
• synthetic-intel
  
• synthetic-ideal
• synthetic-polka
  

2. FCN

github Split network warehouse connections , This warehouse contains FCN, PSPNet, Deeplabv3, Deeplabv3+, DANet, DenseASPP, BiSeNet, EncNet, DUNet, ICNet, ENet, OCNet, CCNet, PSANet, CGNet, ESPNet, LEDNet, DFANet The code of these networks .

We refer to readme Instructions for use , First download pascal_voc Data sets , Then it is used to train the full convolution segmentation network fcn32_vgg16_pascal_voc.

Source code download pascal_voc 2007 and pascal_voc 2012 Data sets , Download pascal_voc 2007 Soon , But download pascal_voc 2012 Very slowly , So I modified the source code , Only download pascal_voc 2007 Data sets , And plan to use pascal_voc 2007 Data sets are trained . because windows Path and linux Path confusion , Debugging failed for many times , So I gave up on windows Training plan on the system . The training will be conducted after the subsequent use of the server .

After training , We can use trained models fcn32_vgg16_pascal_voc Test your image . Because the warehouse does not provide a well trained model , So we can only download the data set step by step , Then train the model , Finally, the model is tested .

3. DenseASPP

github Connect

This code gives the pre training model , I thought the test could be carried out smoothly , But I found that this code is using torch 0.3.1 Written , The version module The name uses '.', such as conv.1, In later versions, module names are not allowed to have '.', So I gave up testing the code .

Now I put my hope on the last choice , Look below .

4.

github Connect

This is a OpenMMLab Do a code warehouse , The warehouse supports splicing different backbone, And support training with different data sets , On this basis , Many segmentation methods are implemented , Based on this code warehouse, we can learn the segmentation network .