AaltoML
are you considering releasing a standalone version that does not rely on Nerfstudio in the future?
Hello, Author! Are you considering releasing a standalone version that does not rely on Nerfstudio in the future? For example, could it be implemented primarily based on the gsplat library for its core functionality?
Hi! Thanks for your interest. Yes, we have a gsplat version with its core functionality, based on gsplat/examples/simple_trainer.py. There are a few differences in the training strategy (e.g., threshold and scene scale), so the results may vary slightly from those reported in the publication. We will release the code in a few weeks.
Hi! Thanks for your interest. Yes, we have a gsplat version with its core functionality, based on
gsplat/examples/simple_trainer.py. There are a few differences in the training strategy (e.g., threshold and scene scale), so the results may vary slightly from those reported in the publication. We will release the code in a few weeks.
Looking forward to it!
Hi! Just wanted to gently check if there's any update on this. Thanks for your time!
Hi! I have uploaded the gsplat version to this repo, please check gsplat branch. The gsplat version is 1.0.0. There are some differences in the pipelines of newer versions, which may not be compatible with our code.
Please note that this is a reproduction, and all results in the paper were obtained using the Nerfstudio codebase. I haven’t adjusted the parameters for the gsplat version, e.g., the thresholds for splitting or cloning and the learning rate. The scene coordinates are also different from the Nerfstudio version.
I reran the experiment on the Yoda scene, and the qualitative results are shown below:
These are the combined images, where the blurred parts indicate the distractors.
The core functions (e.g., splitting and cloning) can be found in examples/simple_trainer_desplat.py.
For dataparsers, please update examples/datasets/colmap.py.
Hope this helps! Feel free to reach out if you have any questions or encounter bugs. Thank you very much for your attention!
Hi! I have uploaded the gsplat version to this repo, please check gsplat branch. The gsplat version is 1.0.0. There are some differences in the pipelines of newer versions, which may not be compatible with our code.
Please note that this is a reproduction, and all results in the paper were obtained using the Nerfstudio codebase. I haven’t adjusted the parameters for the gsplat version, e.g., the thresholds for splitting or cloning and the learning rate. The scene coordinates are also different from the Nerfstudio version.
I reran the experiment on the Yoda scene, and the qualitative results are shown below:
These are the combined images, where the blurred parts indicate the distractors.
The core functions (e.g., splitting and cloning) can be found in
examples/simple_trainer_desplat.py. For dataparsers, please updateexamples/datasets/colmap.py.Hope this helps! Feel free to reach out if you have any questions or encounter bugs. Thank you very much for your attention!
Thanks for that and I will try it