PSMNet test result on Apollo depth dataset

dear JiaRen, I try my best to convert the apollo depth image to the right disparity, it's successful. Then I try to use the disparity to train a model on the PSMNet, and I get some testing result on the testing set, the testing result is not bad, except the sky. The following is the left img, right img, and the testing result. 171206_040521265 171206_040521265-2

the testing command is : python submission.py --maxdisp 208 --model stackhourglass --KITTI 2015 --datapath RR --loadmodel backup/finetune_500.tar

Dec 10 '18 06:12 passion3394

In training, the disparity of sky is all zero. The disparity of the sky in the testing result is near maxdisp(208)

hope to get your advise, thank you very much.

Dec 10 '18 06:12 passion3394

Can't find the dataset you mentioned online. Can you share the link to this dataset? Cloud drive is also OK. https://github.com/JiaRenChang/PSMNet/issues/95#issuecomment-444027415

Dec 11 '18 09:12 forvd

@forvd http://apolloscape.auto/scene.html#to_down_href

Dec 11 '18 09:12 passion3394

Thank you for your help. But i noticed that the apollo dataset doesn't have the depth of movable object. How do you deal with data? Check whether there are moving objects in the picture one by one？

Dec 12 '18 05:12 forvd

@forvd you may use the labelling pics to select the images with little or no movable object.

Dec 12 '18 06:12 passion3394

thank you

Dec 13 '18 01:12 forvd

hello, @passion3394 , the apollo dataset only has depth map,converting it to a disparity map requires the camera's focal length and baseline, but I didn't find these values.　can you help me？　 thank you!