TY - GEN
T1 - Accessible 3D Gaussian Splatting in Google Colab
T2 - 6th International Conference on Innovative Research in Applied Science, Engineering and Technology, IRASET 2026
AU - Cob-Beirute, Daniel
AU - Chavarria-Zamora, Luis
AU - Soto-Quiros, Pablo
N1 - Publisher Copyright:
© 2026 IEEE.
PY - 2026
Y1 - 2026
N2 - Three-dimensional reconstruction underpins applications in robotics, XR, and cultural heritage, yet practical deployment often demands expert setup and high-end hardware. This paper presents a lightweight and reproducible implementation of 3D Gaussian Splatting tailored for Google Colab's free tier (NVIDIA T4). Unlike popular community notebooks that are currently non-functional, our artifact is a maintained, end-to-end notebook that prioritizes stability, simplicity, and accessibility over algorithmic novelty. Concretely, we enable headless reliability for COLMAP, reduce preprocessing time and memory pressure by combining temporal frame subsampling (2 fps) with COLMAP's sequential matching (bounded overlap), configure a free-tier-friendly training schedule (7,000 iterations), and automate packaging of the resulting model for download. The final workflow is linear (cell-based) and requires no additional UI layers, minimizing cold-start overhead and dependency issues in cloud VMs. The contribution is an engineering pathway that makes 3D Gaussian Splatting usable for classrooms, fieldwork, and applied projects, while remaining faithful to established evaluation protocols (PSNR, SSIM, LPIPS) for future quantitative reporting. We release the notebook and instructions to foster adoption and reproducibility.
AB - Three-dimensional reconstruction underpins applications in robotics, XR, and cultural heritage, yet practical deployment often demands expert setup and high-end hardware. This paper presents a lightweight and reproducible implementation of 3D Gaussian Splatting tailored for Google Colab's free tier (NVIDIA T4). Unlike popular community notebooks that are currently non-functional, our artifact is a maintained, end-to-end notebook that prioritizes stability, simplicity, and accessibility over algorithmic novelty. Concretely, we enable headless reliability for COLMAP, reduce preprocessing time and memory pressure by combining temporal frame subsampling (2 fps) with COLMAP's sequential matching (bounded overlap), configure a free-tier-friendly training schedule (7,000 iterations), and automate packaging of the resulting model for download. The final workflow is linear (cell-based) and requires no additional UI layers, minimizing cold-start overhead and dependency issues in cloud VMs. The contribution is an engineering pathway that makes 3D Gaussian Splatting usable for classrooms, fieldwork, and applied projects, while remaining faithful to established evaluation protocols (PSNR, SSIM, LPIPS) for future quantitative reporting. We release the notebook and instructions to foster adoption and reproducibility.
KW - 3D reconstruction
KW - COLMAP
KW - Gaussian Splatting
KW - Google Colab
KW - NVIDIA T4
KW - Structure-from-Motion
KW - accessibility
KW - reproducibility
KW - sequential matching
UR - https://www.scopus.com/pages/publications/105042513645
U2 - 10.1109/IRASET68627.2026.11538846
DO - 10.1109/IRASET68627.2026.11538846
M3 - Contribución a la conferencia
AN - SCOPUS:105042513645
T3 - 2026 6th International Conference on Innovative Research in Applied Science, Engineering and Technology, IRASET 2026
BT - 2026 6th International Conference on Innovative Research in Applied Science, Engineering and Technology, IRASET 2026
A2 - Benhala, Bachir
A2 - Raihani, Abdelhadi
A2 - Qbadou, Mohammed
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 14 May 2026 through 15 May 2026
ER -