Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Evaluation give different score from the paper #9

Open
pureexe opened this issue Nov 6, 2023 · 1 comment
Open

Evaluation give different score from the paper #9

pureexe opened this issue Nov 6, 2023 · 1 comment

Comments

@pureexe
Copy link

pureexe commented Nov 6, 2023

I would like to thank you for creating such a great work.

I'm currently working on light estimation and want to compare it with the Stylelight.

However, when I ran your code on the Laval indoor dataset, the score is quite different. So, I would like to ask if this is still acceptable. If not, I am looking for the correct way to run an evaluation.

Report in paper: M (Mirror ball) Run by myself: M (Mirror ball)
Angular Error 4.30 6.74
RMSE 0.56 0.58
si-RMSE 0.55 0.56
Report in paper: S (Silver matte ball) Run by myself: S (Silver matte)
Angular Error 2.96 4.61
RMSE 0.30 0.33
si-RMSE 0.29 0.32
Report in paper: D (Diffuse ball) Run by myself: D (Diffuse ball)
Angular Error 2.41 4.07
RMSE 0.15 0.16
si-RMSE 0.11 0.13

Let me specify how I got this score so you can point out which step I did wrong.

  1. Prepare the cropped Laval indoor dataset by running data_prepare_laval.py
  2. changing input path by pointing root_path in test_lighting.py to the corrected path
  3. Tonemap the ground truth in the directory name test using evaluation/tonemap.py
  4. Tonemap the output of Stylelight using evaluation/tonemap.py
  5. Render ground truth into 3 balls using evaluation/test_render.sh
  6. Render the output of Stylelight into 3 balls using evaluation/test_render.sh
  7. Compute the score by run evaluation/test_rmse.sh --fake <stylelight's tone-mapped dir> --real <ground truth's tone-mapped dir>

Best regard.

@Wanggcong
Copy link
Owner

Sorry, I am very busy these days. I am afraid that I have no time to run the code and check the differences. One thing I think you might miss is that I removed the black regions and warped the panoramas to some extent such that the object is placed at a new center. This is similar to the method Gardner et al. Learning to Predict Indoor Illumination from a Single Image, SIGGRAPH Asia, 2017..

You might keep the same setting for fair comparison when compared with other methods, since I found lots of hyperparameters would affect the final results.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants