Open Ujjwal-Rohilla opened 1 week ago
Try to read someting about how to take 'best' pictures for colmap. In general you need study 'photogrametry' with the twist ;) Here is nice tutorial https://radiancefields.com/gaussian-splatting-brings-art-exhibitions-online-with-yulei
GL
Hi, thanks for the reply
I read through it and (also before this read the colmap docs one more time) and understood one main thing, which is that colmap prefers translations more than rotations (which I guess makes sense now I think about it)..
Also, I re-recorded my apt hallway with only x y z movement, and also got the latest GUI version of colmap and tried it and it worked out well! Excited for recording the staircase and running that too thru the colmap.. Thanks for the help!!
Here's the final result of the gsplat btw (I only took ~35 images so ye)
Now - hard part :) Make it good in the both ways ;)
Ahhh I see what you mean lol
a 360 camera would be best for this but the best I have is a nikon camera and my phone... if I walk in the hallway and rotate near the end and then walk back... that'll work? I'ma try it out
Did you use convert.py
to preprocess your images to build COLMAP priors? I am not sure if it helps, but it worked well for me.
I have been trying to generate a Gaussian splat of my school's staircase, however it is failing at the COLMAP part. And the fact that it takes some time for it to process I decided to use a smaller dataset of my apartment's hallway.
The video is completely stable and stuff and using ffmpeg I got 613 Images. Running COLMAP goes as expected, however in the end it just says "Undistorting Image [1/2]" and "Undistorting Image [2/2]" for whatever processing it needs to do, which generates poor Gaussian splats in the end.
Here is my dataset
Also, the "images" that the COLMAP produces are like very distorted, as if the undistorting didnt work (or overworked idk)
Any help would be appreciated