Optimized Minimal 3D Gaussian Splatting

🔼 This figure illustrates the architecture of the Optimized Minimal Gaussians (OMG) model. The OMG model efficiently represents 3D scenes using a minimal number of Gaussian primitives. It learns both geometric (position, scale, rotation) and appearance (static and view-dependent color, opacity) features for each Gaussian. These features are then compressed using Sub-Vector Quantization (SVQ) to reduce storage. The geometric attributes (after SVQ) are used directly for rendering. A novel spatial feature, derived from the Gaussian’s position, is incorporated into the appearance features to improve rendering quality, particularly in areas with sparse Gaussians. This combination balances the need for compact representation with accurate rendering.

read the caption

Figure 2: The overall architecture of our proposed OMG. OMG learns per-Gaussian geometric and appearance features, applying Sub-Vector Quantization (SVQ) to all of them. The SVQ-applied geometric attributes are used for rendering, while the space feature based on the Gaussian center position is integrated into the appearance features to define the final appearance.

🔼 This figure illustrates three different vector quantization methods used for compressing Gaussian attributes in 3D Gaussian splatting. (a) shows standard vector quantization, where the entire attribute vector is encoded using a single codebook. (b) depicts residual vector quantization, which involves multiple stages of encoding where the residuals from previous stages are encoded. (c) presents sub-vector quantization, which partitions the attribute vector into multiple sub-vectors and uses separate codebooks for each, reducing the computational complexity of large codebooks while maintaining precision. The ‘+’ symbol represents element-wise summation, and the ‘⊕’ symbol denotes vector concatenation.

read the caption

Figure 3: Conceptual diagram of (a) vector quantization, (b) residual vector quantization, and (c) sub-vector quantization. + and ⊕direct-sum\oplus⊕ denote the element-wise summation and the vector concatenation.

🔼 Table 2 presents a comparison of the performance of the Optimized Minimal Gaussians (OMG) method against several baseline methods on two datasets: Tanks & Temples and Deep Blending. The metrics used for comparison include PSNR, SSIM, LPIPS, size (in MB), and FPS. For each method, the table shows the results obtained using an NVIDIA RTX 3090 GPU, which is consistent with the LocoGS [50] paper used as a baseline. Additionally, to provide a more comprehensive evaluation, the table also includes performance results measured using a higher-performance NVIDIA RTX 4090 GPU, which is indicated by values in parentheses. This allows for better understanding of the relative performance gains across different hardware.

read the caption

Table 2: Quantitative results of OMG evaluated on the Tanks&Temples and Deep Blending datasets. Baseline results are sourced from the LocoGS [50] paper, where the rendering results were obtained using an NVIDIA RTX 3090 GPU. Our rendering performance was measured using the same GPU, with the values in parentheses obtained from an NVIDIA RTX 4090 GPU.

🔼 This table compares the efficiency of different variants of the Optimized Minimal Gaussians (OMG) representation with the state-of-the-art LocoGS method on the Mip-NeRF 360 dataset. It shows a comparison of training time, the number of Gaussian primitives used, the resulting storage size, and the rendering quality (PSNR, SSIM, LPIPS) for each method. This allows for a quantitative assessment of the trade-offs between model complexity, storage efficiency, and rendering quality.

read the caption

Table 3: Efficiency comparison of OMG variants compared to LocoGS evaluated on the Mip-NeRF 360 dataset. We present training time, the number of Gaussians, and the storage requirement with rendering quality.

🔼 This table presents an ablation study on the Optimized Minimal Gaussians (OMG) model, specifically examining the impact of two key components: the space feature integration and the local distinctiveness (LD) scoring. It uses the Mip-NeRF 360 dataset for evaluation and shows the PSNR, SSIM, LPIPS metrics, the number of Gaussians, and the model size for different configurations of OMG. The configurations vary based on the inclusion or exclusion of the space feature and LD scoring, allowing for a quantitative assessment of their individual contributions to the overall performance of the OMG model.

read the caption

Table 4: Ablation study of OMG using the Mip-NeRF 360 dataset. We evaluate the contribution of the space feature integration and local distinctiveness (LD) scoring.

🔼 This ablation study analyzes the impact of using Sub-Vector Quantization (SVQ) versus Vector Quantization (VQ) in the Optimized Minimal Gaussians (OMG) model. The experiment focuses on the Mip-NeRF 360 dataset. The table compares performance metrics (Training time, number of Gaussians, model Size, PSNR, SSIM, LPIPS) for OMG-XS with SVQ and the same model but with SVQ replaced by VQ. This allows for a direct comparison of the two quantization methods and reveals the effects on model performance and training efficiency.

read the caption

Table 5: Ablation study on SVQ using the Mip-NeRF 360 dataset. We substitute SVQ to VQ.

🔼 This table presents an ablation study analyzing the impact of different post-processing techniques on the overall size of the Optimized Minimal Gaussians (OMG) model. It shows the model sizes resulting from using various combinations of G-PCC compression, Huffman coding, and LZMA compression. This allows assessment of the effectiveness of each compression method independently and in combination.

read the caption

Table 6: Ablation study on the post-processing methods applied in OMG.

🔼 This table details the average storage used by each component (position, scale, rotation, appearance features, and MLPs) within the Optimized Minimal Gaussians (OMG) model. Different versions of the OMG model (XS, S, M, L, XL) are compared, showing how storage needs vary with model complexity. The ‘Actual size’ column represents the total file size for each model variant, encompassing all components.

read the caption

Table 7: The average storage allocation for each component across OMG variants. ‘Actual size’ refers to the total size of a single file containing all components.

🔼 This table presents a comprehensive breakdown of the per-scene performance metrics obtained from evaluating the Optimized Minimal Gaussians (OMG) model on the Mip-NeRF 360 dataset. It details the PSNR, SSIM, and LPIPS values for various scenes (bicycle, bonsai, counter, flowers, garden, kitchen, room, stump, treehill) across different configurations of the OMG model (OMG-XS, OMG-S, OMG-M, OMG-L, OMG-XL). In addition to the image quality metrics, the table includes the training time, the number of Gaussians used, the model size (in MB), and the rendering FPS for each scene and model variant. This allows for a detailed comparison of the trade-offs between model size, computational cost, and visual fidelity across different scenes and OMG configurations.

read the caption

Table 8: Per-scene results evaluated on the Mip-NeRF 360 [2] dataset.

🔼 This table presents a detailed breakdown of the performance of the Optimized Minimal Gaussians (OMG) model on a per-scene basis for two datasets: Tanks & Temples and Deep Blending. For each scene within each dataset, the table provides key metrics such as PSNR, SSIM, and LPIPS, offering a comprehensive evaluation of the visual quality achieved by the model. Additionally, the training time, the number of Gaussians utilized, the overall file size, and the frames per second (FPS) achieved during rendering are reported, providing a complete picture of both the model’s accuracy and efficiency.

read the caption

Table 9: Per-scene results evaluated on the Tank&Temples [28] and Deep Blending [21] datasets.