search

prs-eth / Marigold

[CVPR 2024 - Oral, Best Paper Award Candidate] Marigold: Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation
https://marigoldmonodepth.github.io
Apache License 2.0
2.42k stars 132 forks source link

issue with reproducing results in the paper #96

Closed yongmayer closed 4 months ago

yongmayer commented 4 months ago

Hi,

I try to reproduce the results in the paper. The result looks good for NYU-v2, but for KITTI dataset, it didn't give a close match with the paper.

NYU-v2

abs_relative_difference 5.35 delta1_acc 97.38

KITTI abs_relative_difference 12.28 delta1_acc 86.58

Wondering if the smaller datasets or the full datasets (commented out) in dataset_val.yaml are used in the paper for validation? The paper mentioned "Eigen test split [ 14 ] made of 652 images" for validation on KITTI dataset, but I couldn't find the corresponding data split file. Would you mind sharing it?

Thank you very much for any help!

markkua commented 4 months ago

Hi, the full datasets are used for the paper evaluation. You can find the corresponding config files in the scripts here. As for KITTI dataset, this data split is the one you are looking for. There are some files with out GT, resulting in 652 valid samples at the end.