Reproduce “Trick-GS: A Balanced Bag of Tricks for Efficient Gaussian Splatting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Anil Armagan, Albert Saà-Garriga, Bruno Manganelli, Mateusz Nowak, Mehmet Kerim Yucel
Trick-GS combines training strategies to reduce the storage and training burden of 3D Gaussian Splatting while retaining its representation. It prunes low-impact Gaussians and color attributes, progressively trains from blurred, lower-resolution images, and reduces repeated computation. On Mip-NeRF 360, it uses 39 MB versus 770 MB and renders at 222 versus 142 FPS, with PSNR of 27.16 versus 27.56 dB on an RTX 3090.
Key insights
- A systematic combination of diverse optimization strategies can yield comprehensive efficiency improvements across model size, training speed, and rendering speed for 3D Gaussian Splatting, without significant quality compromise.
- Progressive training techniques effectively regularize the scene representation learning, leading to better initialization, faster convergence, and improved structural consistency in reconstructed scenes.
- Strategic pruning and masking of Gaussians and their view-dependent attributes are crucial for drastic model size reduction while maintaining high visual fidelity, making the technology viable for edge devices.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.