Results from code and hf demo is different for simple_radial model

[bbzh]

Hi,

I observe that the result by running the code is generally worse than the one in hf demo for the simple_radial model.
As an example, the first one below is from hf while the second one is from running code.
I can get the same results for the simple_divisional model, though.
Can you check if the code is using the same parameters as in hf demo? Thank!

[veichta]

Thank you for pointing this out! Are you running the demo notebook? It uses weights trained on distorted images, while the Hugging Face demo always uses weights trained on pinhole images. This mismatch likely explains the difference in results for the simple_radial model.

For this image, the simple_divisional model is more appropriate, as it handles stronger distortions better. Additionally, for the simple_radial model, the estimated focal length dictates how strong of a distortion can be handled, which might be a limitation in this case.