Sub-pixel offsets in quickview images introduce "fringe" artifacts

[dshean]

We noticed these sub-pixel artifacts during initial tests with the quickview images. They are not present when we do the downsampling of the full-res images with gaussian pyramid approach.

They were likely introduced during the resampling, and perhaps related to odd pixel dimensions when downscaling...

hipp/src/hipp/image.py

Line 80 in 961955a

interpolation: int = cv2.INTER_LINEAR,

I'm a little surprised, as I would hope that bilinear would be sufficient. Suggest using cubic (or ideally gaussian) when creating downsampled products in the future, and/or more carefully specifying subsampling factor and dimensions (as mentioned in #7).

Also important to specify downsampling factor of 20 was used to prepare the KH9 PC quickview images somewhere in the data document.

[godinlu]

In practice, I had not anticipated that quickview images could be used for preprocessing purposes.
This is why I initially opted for a simple interpolation method and the JPEG format, which inherently reduces image quality.
As for the resizing factor, it is currently specified only at the beginning of each notebook, under the settings section: QUICKVIEWS_FACTOR.

For preliminary processing on downscaled images, I can, however, generate a new dataset using a more advanced interpolation method, such as bicubic interpolation, while preserving the TIFF format. I would also ensure that the same scaling factor is applied consistently across all the data.

[dshean]

All sounds good.

Upon further review, we're still seeing some of these fringe artifacts at coarse resolution images prepared from the full-res preprocessed images, though they're much less noticeable.

I know we've looked at this in the past (including earlier discussion with Amaury), and it may be limitation of UInt8 images with limited contrast. Several potential steps in the scanning, stitching, and preprocessing workflows where this could be introduced. Will require more careful analysis, but not a priority right now.

[adehecq]

I have noticed these fringe patterns many times with the KH-9 MC images in ASP, even at full resolution. They occur essentially over very flat terrain. In my case, over ocean surfaces with some sea ice that can be matched during the dense matching. But Casa Grande is also very flat so I am not surprised.
I am not sure they are introduced during the preprocessing. Could be also a rounding issue somewhere during the stereo processing.