the prediction on in-the-wild example is noisy

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

2.39k stars 132 forks source link

the prediction on in-the-wild example is noisy #76

Closed indu1ge closed 4 months ago

indu1ge commented 5 months ago

Hi, I trained the model with your official training code. The quantitive results are consistent with the paper. But the qualitive result is relatively poor with much noise. Could you provide some suggestions on that? Screenshot_2024_0603_160949

markkua commented 5 months ago

How many iterations have you trained? I would recommend training further, as qualitative results might still be improved although one might not observe in the quantitative results.