Bad prediction quality out of the box (wrong config?)
isarandi opened this issue · comments
Thanks for releasing this work!
I'm trying to make predictions on images using inference.py, but the results look very misaligned compared to the image data. I didn't change anything, other than the fixes mentioned in the FAQ to get it running.
This is the output as drawn by this codebase:
And I get the same if I project the saved obj file myself using the saved intrinsics json data:
Based on the quantitative evaluation results, the output should be much better aligned, I wonder if something is configured wrong.
The HuggingFace demo gives the same result, so I think this is the real result.