I took 3000 pictures of my tool cart yesterday. I was rotating three spare batteries to keep the camera alive. I didn’t use any of the pictures cause I was too slow & the lighting was inconsistent throughout. I also bumped into the cart halfway thru the process. About to go out and do it all over again!
How do you process the images?
Every step that comes after I put the images in a folder is pretty much entirely "agentic". I rent a suitable 4090 GPU from vast.ai (costs around 50 cents/hr), pass the ssh info to my agent, and let it take care of the COLMAP reconstruction + training processes. My current task is to transform this pipeline into something much more deterministic so that I don't have to burn tokens to train the next splat.
Wow, that’s very involved. For some reason I expected you’d be doing it on a home machine. How long does it take on the rented GPU?
reply
add comment