search

xy-guo / Learning-Monocular-Depth-by-Stereo

Learning Monocular Depth by Distilling Cross-domain Stereo Networks, ECCV18
https://arxiv.org/abs/1808.06586
MIT License
94 stars 8 forks source link

Ground Truth Depth Map #11

Closed DeepDeep123 closed 5 years ago

DeepDeep123 commented 5 years ago

Hi, Guo,

I am wondering which ground truth depth map you use for supervised fine-tuning and evaluation on KITTI dataset. Is the offical depth provided by KITTI or from monodepth. I am confused, which one shall we use?

Thanks!

xy-guo commented 5 years ago

The official depth provided by RAW KITTI is in point cloud format, which cannot be directly used for training. The code from Monodepth projects the point cloud into a depth map to the left camera view. My code follows Monodepth and also projects the point cloud into a depth map in the left camera view. The only difference is that Monodepth utilizes LiDAR depth values, while I use the depth values in the camera view. You can check this file for the depth map generation.

DeepDeep123 commented 5 years ago

Hi, Guo,

Thanks for your reply! I just find that KITTI now indeeds provide ground truth depth map data_depth_annotated.zip, where ground truth maps are saved as 16bit PNG files.

By the way, you use depth values in the camera view, does it mean that vel_depth is set to be False?

Thanks!

xy-guo commented 5 years ago

I think the data_depth_annotated.zip is from the new KITTI Depth Prediction benchmark, while my paper is based on KITTI RAW data and the Eigen split.

It is the same when you set vel_depth to False.