Closed jhkennedy closed 10 months ago
@alex-s-gardner I did confirm we'll want these changes in production. Ideally, we'd cut a v1.5.1 release with these changes, and then hyp3-autorift can use the "canonical" version from conda-forge. I can move faster than that, however, with just these changes.
I suggest reviewing the code related to this operator and implementing a more robust check to prevent negative variances. Additionally, ensure that the valid data is handled correctly, avoiding zero valid pixels. This fix will enhance the stability and reliability of the wallis_fill filter.
Feel free to share any specific code snippets or files related to this issue, and we can work together to make the necessary adjustments. Collaboration is key in open source, and your contribution in identifying this problem is greatly appreciated. Let's work together to resolve it and improve the project's overall quality.
I suggest reviewing the code related to this operator and implementing a more robust check to prevent negative variances. Additionally, ensure that the valid data is handled correctly, avoiding zero valid pixels.
@itstalmeez I'm happy to review a pull request if you do any of the above or have further suggestions.
wallis_fill
filtering images, the variance can end up less than zero (but very close to zero), leading to NaNs in the valid data, which blows up downstream and result in zero valid pixelswallis_fill
filtering images