search

ifnspaml / SGDepth

[ECCV 2020] Self-Supervised Monocular Depth Estimation: Solving the Dynamic Object Problem by Semantic Guidance
MIT License
200 stars 26 forks source link

KITTI zhou split has different number of training images from Monodepth2 #29

Open songlin opened 2 years ago

songlin commented 2 years ago

Hi,

Thanks for sharing the awesome work. I use the kitti_zhou_split training split provided by you here. It contains 21880 training images from each left and right camera. But the same zhou split from Monodepth2 contains 19905 imageshere.

Can you explain why? or point me the mistakes i have.

klingner commented 2 years ago

Hi,

the original split from [Zhou et al., SfMLearner, CVPR 2017] contains 40,109 images, the split from [Godard et al., monodepth2, ICCV 2019] contains a subset of 39,100 image triplets. In the code I provide I basically use all left and right images, where at least one image (left or right) exists in Godard et al.'s split. This results in the number of image triplets you find here in this code.