What are the tradeoffs with re-sizing images?

I’d like to understand what I’m giving up by allowing the processing pipeline to resize images. I have a hard time imaging why that doesn’t degrade the outputs, or whether some outputs are affected more than others by resizing. Thanks… Bob R.

1 Like

You are (potentially) losing quality, but not always and probably not as much as you think. The reason is that the final orthophoto’s resolution often times does not require full size images (think about it, if you have 100 images with 4000x3000 but your orthophoto image size is 25000x10000 you know that you haven’t used all the pixels at your disposal). Similarly the point cloud is already calculated with a resized version of the input images (default is 640 pixels), so unless you go lower than that, you won’t notice significant decrease in quality in the 3D output.

Of course full resolution images give you better feature extraction and matching, so the reconstruction quality might improve overall, texturing will also look sharper due to the lack of interpolation (which you get when you resize images) and the gradient magnitude score being affected during texturing (look up the mvs-texturing paper for details on that), but often times the memory and time savings of resizing the images are worth it, especially if the results look good enough.

It’s always a tradeoff between speed and quality and ODM tries to find a good balance.

3 Likes

Excellent explanation. Exactly what I needed. Thanks… Bob R.

1 Like

Sorry, Why 2048px is default of resize images?

1 Like