Came across some photogrammetry software lately. I had an idea to try them out by taking a bunch of photos and generate a 3d model.
But it’s difficult. Tools I tried were COLMAP and visualSFM. Seems to be bad results. I think I had insufficient number of photos, and there weren’t enough correlated features.
I thought “well what if I used a video”, so I took a video of my mouse, and dump them into images.
But a lot of them were blurry, like this.
So I thought there should be a way to filter out the blurry images. Such as using
- Fast Fourier Transform
- Laplace (or LoG) filter
There were some suggestions on StackOverflow that hint OpenCV is a good tool. I found some blurry detector on github https://github.com/indyka/blur-detection but there was some issues. For one, the threshold computed varied too greatly and so I had to manually figure out what were the 80th percentile and better (sharper, higher score) images. But some images contained no content, such as a blank table. It would be classified as sharp… so if going down this method, you might want to remove those blank images.