ELMGS: Enhancing memory and computation scaLability through coMpression for 3D Gaussian Splatting

TL;DR

This work proposes an iterative pruning strategy that removes redundant information encoded in the model and enhances compressibility for the model by including a differentiable quantization and entropy coding estimator in the optimization strategy.

Abstract

3D models have recently been popularized by the potentiality of end-to-end training offered first by Neural Radiance Fields and most recently by 3D Gaussian Splatting models. The latter has the big advantage of naturally providing fast training convergence and high editability. However, as the research around these is still in its infancy, there is still a gap in the literature regarding the model's scalability. In this work, we propose an approach enabling both memory and computation scalability of such models. More specifically, we propose an iterative pruning strategy that removes redundant information encoded in the model. We also enhance compressibility for the model by including in the optimization strategy a differentiable quantization and entropy coding estimator. Our results on popular benchmarks showcase the effectiveness of the proposed approach and open the road to the broad deployability of such a solution even on resource-constrained devices.

Figures (6)

• Figure 1: Qualitative comparison of ELMGS with 3DGS kerbl20233d, C3DGS niedermayr2023compressed, Compact3DGS lee2023compact, and Reduced-3DGS papantonakis2024reducing. With our proposed method we can achieve compression rates of about 27$\times$ with indiscernible loss in visual quality and significantly better rendering speed as compared to other methods.
• Figure 2: ELMGS begins with a pre-trained 3DGS scene and performs iterative pruning with finetuning to remove less significant Gaussians, followed by Quantization-Aware finetuning. The quantized model is then entropy-encoded to generate the final compressed scene.
• Figure 3: Effect of gradual pruning to the 3DGS model: a gradual removal allows the model to self-adjust and to better fit the scene/object.
• Figure 4: Effect of gradual pruning to the opacity values. We have two effects: (i) the number of parameters in our model reduces; (ii) the frequency around specific low and high density is increased.
• Figure 5: Comparison of ground truth images from the test set of bicycle, drjohnson, and truck scenes between ELMGS "ours-big", "ours-medium" and "ours-small" compressed representation and 3DGS-30k.