search

JiaRenChang / PSMNet

Pyramid Stereo Matching Network (CVPR2018)
MIT License
1.41k stars 424 forks source link

How can I test my own dataset? #38

Open blueevangelion opened 6 years ago

blueevangelion commented 6 years ago

I tried test some of my own stereo images. I got error like ValueError: cannot reshape array of size 1863000 into shape (1,3,375,1242).

What should i do? Is that a easy way to test a pair of stereo images?

JiaRenChang commented 6 years ago

@blueevangelion

Hi, You can modify the predefined image width and height in the code.

blueevangelion commented 6 years ago

I run into out of memory problem while I running the training script on 1080Ti(3 of them). Any suggestion?

kingsleyluoxin commented 6 years ago

You can reduce the size of batches, then it will be solved.

dongdongcomeon commented 6 years ago

Hi,thanks for your working!I’am a Chinese student.Recently,I apply railway images for stereo matching and plan to use your pretrained model to test my own railway images.However,I am afraid I can‘t get the expected result.Can you give me suggestion for stereo matching?This is my railway iamges!Thanks! 11

JiaRenChang commented 6 years ago

@dongdongcomeon Hi, I do not know the setting of your images. In our pretrained model, the maximum disparity is 192. You may consider this constraint to apply your images.

dongdongcomeon commented 6 years ago 
   您好！感谢您的回复！我的影像最大视差为480，您说您所使用的数据集最大视差为192，那么我的数据是否还适用于您的模型呢？
    我的影像大小为3840*5760，比您论文中使用的影像的大小要大，这使用您的模型时，会有什么影响吗？
JiaRenChang commented 6 years ago

@dongdongcomeon 您好，我的建議是您可以將影像downsample到1/4大小，這樣視差也跟著下降到1/4

passion3394 commented 5 years ago

@JiaRenChang 是宽和高都缩放到原来的1/4吗？还是都缩放到原来的1/2？