PLEASE SHARE THIS TO INCREASE ITS REACH
I have a challenge for
#AI and
#machinelearnig engineers,
#Python #TensorFlow #Karas and
#OpenCV developers:
I bet you can't develop an app that crops and straightens
#filmstrip frame scans in the way I'm doing it by hand in this video.
Each frame is overscanned to ensure I get the whole thing. Because "real" 35mm scanners cost $1M or more, I am using a flatbed scanner to save this
#lostmedia. The position of the frame in each image may be different, and each frame may be skewed up to 0.4 degrees in either direction.
A small amount of overscan needs to remain around the frame, because the frame mattes used to make these films were never precisely square. Each frame should be straightened relative to its left edge, and then the image cropped down to just outside the edge of the frame. Additionally, each resultant image from a particular film should be the same resolution, with the frame image centered in the crop.
Apparently this is beyond the capabilities of every software developer on earth, and outside of the understanding of every AI and non-AI app currently available.
Should you accept this challenge, you will be helping a single independent archivist trying to save an entire film format, almost all titles of which are currently lost media. The scans need to be reassembled into videos. These videos will be preserved at the
@internetarchive. If you were to develop a solution that works, I would be happy to train it by continuing to manually do this until it learns to do it by itself.
(Additionally, it would be interesting to hear from developers as to why this sort of thing is impossible when ChatGPT can pass the bar exam and Grok can generate photos of Donald Trump riding a velociraptor. I can't help but think what I'm asking is orders of magnitude simpler than either of those things..)