Repository navigation
Killed when the data maybe too big #134
Description
I don't know if I've tested it, I used a large-scale data (colmap), put it into OpenSplat, nerfstudio's nerfacto and 3d_gaussian_splatting respectively, and only OpenSplat was killed when importing, nerfstudio trained normally, which factor formed such a difference? Will a fix be attempted in the future?
Activity
pierotofy commentedon Sep 27, 2024
We'd welcome a pull request that fixes/optimizes any memory issue that might be present in OpenSplat. 🙏
mureva commentedon Oct 4, 2024
I don't know if relates to your problem, but I was trying to use a lidar point cloud as the initial point cloud for splatting, and would get a runtime errror. A bit of sleuthing suggested that the issue was nanoflann, an it appeared to be getting to the recursion limit of the system. An older version of nanoflann (1.3.0) was able to deal with my point cloud without problem - but meant a small amount of modification to opensplat to use.
liamyehudai commentedon Oct 8, 2024
We are also having this issue
psykon commentedon Oct 19, 2024
Same issue here. I have a colmap dataset with 3564 images. OpenSplat seems to attempt to load all images into memory at once. After committing ~160Gb to memory, the process shuts down with a cv::OutOfMemoryError message.
This happens with the prebuilt windows binary.
Switching to CPU instead of cuda does not make any difference, neither does setting a downscale factor make a difference in this behavior.
Is this a expected limitation of OpenSplat or a unexpected error?
pierotofy commentedon Oct 19, 2024
You need more RAM, currently. We'd welcome a PR that improves memory management.
SubodhMishraIdeaForgeTech commentedon Jan 29, 2025
I downsampled the images and now it's running. Is this a possible way to address the problem?
pierotofy commentedon Jan 29, 2025
A way to improve this would be to lazy-load images and offload them as needed (via a LRU cache structure) in between training iterations, which would slow down training but reduce RAM usage. We'd welcome improvements in this area.
luisseyfer commentedon Jan 31, 2025
@SubodhMishraIdeaForgeTech
Could you provide some information?
What is your hardware? What was the data size first? How much did you downsample?
With some examples one could maybe estimate...
CarCrasher13 commentedon May 17, 2026
this is needed.
pierotofy commentedon Sep 16, 2026
#250 should allow processing of much larger datasets by using a compressed memory cache. 🥳