How to determine winsize (key images) and refid (neighbors) in mast3r_sfm?

[Reed-yang]

I want to commend the use of a Retrieval model to connect views in mast3r_sfm—it's a very smart approach and has worked well in my current image testing.

However, I’ve noticed that in the demo code, the winsize (key images) and refid (neighbors) parameters are fixed and don’t scale with the number of images. For example, even when processing several hundred images (in large-scale scene image sequences), these parameters remain capped at 20 and 10, respectively, which is the same as when working with just a few dozen images.

Here are my questions:

• For large-scale scenes, do these parameters need to be adjusted manually? Based on your experience, what would be the best way to calculate winsize and refid (depending on num_files) to balance both the final quality and the number of image_pairs effectively?

[michaelyuancb]

@Reed-yang Have you ever checkout the dense pointclouds quality ? I found the pointclouds split into "patchs" in my setting. I wonder if this is normal phenomenon for Master-Sfm winsize strategy ?

[Reed-yang]

@michaelyuancb split into "patchs" may have relations with the anchor depth mentioned in the paper, by default, the the num of anchors should be H x W / 64