About matching time comlexity

[LaFeuilleMorte]

Hi, Thanks for open-sourcing this brilliant work! I've tried reconstruction on a bunch of about 200 images. However my card crashes with the OOM errors. So I tried the "swin" (sliding window) option. But the results was way worse than using "comlete" option. Is there any trade-off measures I can use to achieve balance between performance and computational complexity? Thanks for your reply!

[Reed-yang]

What type of aligner did you use? I was suffering OOM as well using the dust3r global_aligner , and, the sparse_global_alignment used cache strategy solves this.

[LaFeuilleMorte]

What type of aligner did you use? I was suffering OOM as well using the dust3r global_aligner , and, the sparse_global_alignment used cache strategy solves this.

I m using spare_global_alignment. The oom error happens when too many frames were input (over 200 images, which will create 38000 pairs if using "complete" option). Mine solution is to seek a way to avoid handling 38,000 pairs. When I was using "swin-3" option. The time complexity was 3*n instead of n^2. Which will significantly lowers the computational source and time. But I have no idea which window length to choose to achieve a best balance between performance and computation comlexity.

[Reed-yang]

if your data have no explicit pattern, such as timestamps or ego centric, using logwin maybe better https://github.com/naver/mast3r/issues/21#issuecomment-2250587429 the key is to utilize your data's inner connections

btw, your OOM means CUDA OOM? I was thought to be RAM OOM

[LaFeuilleMorte]

if your data have no explicit pattern, such as timestamps or ego centric, using logwin maybe better #21 (comment) the key is to utilize your data's inner connections

btw, your OOM means CUDA OOM? I was thought to be RAM OOM

Oh I guess it's RAM oom coz I received a "killed" message.

[eyildiz-ugoe]

this was an issue with dust3r and it still is an issue with mast3r. cool names, but poor performance. it collapses after a handful of images, same as dust3r. they failed to program it in accordance to available RAM. i just tried mast3r with the hope that they actually may have fixed that but no, sad.