Reconstruct-It! A Collection of 3D Reconstruction Datasets and Trained Splats Marginally Better blog

The data required for training NeRFs and Gaussian Splats is often much smaller than what used to be required for doing 3D reconstruction. There are quite quite a few formats for the per-scene datasets for training these methods. I often had problems on training such radiance fields on multiple datasets while working on any training, thus I share a collection of 3D reconstruction scenes gathered from various sources, each accompanied by a trained Gaussian Splat.

There are 77 scenes in this collection. Each scene in the dataset usually includes:

Multi-view image sequences with camera parameters
Calibration information
A pre-trained Gaussian Splat model optimized for the scene

All the datasets in the collection are taken from various different sources and have different licenses.

Citation

Please cite this work as:

Rishit Dagli. "Reconstruct-It! A Collection of 3D Reconstruction Datasets and Trained Splats". Rishit Dagli's Blog (March 2025). https://rishit-dagli.github.io/2025/03/28/nerf-gs-datasets.html

Or use the BibTex citation:

@article{ dagli2025reconstruct-it!,
  title = { Reconstruct-It! A Collection of 3D Reconstruction Datasets and Trained Splats },
  author = { Rishit Dagli },
  journal = {rishit-dagli.github.io},
  year = { 2025 },
  month = { March },
  url = "https://rishit-dagli.github.io/2025/03/28/nerf-gs-datasets.html"
}

Citation

Tags