Asking about abnormal predicted depth maps

zzangjinsun / NLSPN_ECCV20

Park et al., Non-Local Spatial Propagation Network for Depth Completion, ECCV, 2020

MIT License

321 stars 57 forks source link

Asking about abnormal predicted depth maps #12

Closed HuynhLam closed 3 years ago

HuynhLam commented 3 years ago

Hi, This is such a great work, with impressing results! I've tried to generate the depth maps using your pretrained model but somehow the depth image doesn't look good (Inputs from KITTI validation selection set). 2011_09_26_drive_0002_sync_image_0000000005_image_02 Predicted depth map 0000000 The only difference with yours recommended setting is that I don't have x4 GTX GPUs. Can this be the reason that affect the final depth map? Thank you.

zzangjinsun commented 3 years ago

Hi @HuynhLam,

Did you use the latest code and the command I provided?

For the KITTI dataset with the pretrained model, the following arguments are important:

--max_depth 90.0 --num_sample 0 --preserve_input

Otherwise, could you let me know your input arguments? (You can find args.json in the saving directory)

HuynhLam commented 3 years ago

Hi @zzangjinsun, Thanks for your fast response. I missed the _--preserveinput argument. It works fine now. Thanks a lot! I am gonna close this issue.