Mini-Splatting: Representing Scenes with a Constrained Number of Gaussians This paper introduces Mini-Splatting, a method for efficiently representing scenes using a constrained number of Gaussians. The key challenge is to maintain high-quality rendering while reducing the number of Gaussians. The authors analyze the inefficiencies of Gaussian representation in 3D Gaussian Splatting (3DGS), where Gaussians are often clustered, leading to suboptimal rendering quality and speed. To address this, they propose a densification and simplification algorithm that reorganizes the spatial positions of Gaussians, resulting in a more uniform distribution and improved performance. The densification process includes blur split and depth reinitialization. Blur split involves splitting Gaussians in areas with smooth color transitions, while depth reinitialization uses merged depth points to reinitialize the scene. The simplification process includes intersection preserving, which retains Gaussians with the most impact on the rendered image, and sampling, which preserves both geometric structure and rendering quality. The proposed methods are integrated into the original rasterization pipeline, providing a strong baseline for future research in Gaussian-Splatting-based works. The results show that Mini-Splatting achieves a balanced trade-off between rendering quality, resource consumption, and storage. The method is evaluated on multiple benchmarks and datasets, demonstrating its potential and scalability. The authors also discuss related work, including Gaussian Splatting and 3D data simplification. They highlight the challenges of Gaussian simplification, such as inefficient spatial distribution and the need for effective pruning strategies. The paper presents ablation studies and quantitative results, showing that their methods outperform existing approaches in terms of rendering quality and efficiency. The paper concludes that Mini-Splatting effectively represents scenes with a constrained number of Gaussians, achieving high-quality rendering while reducing computational and storage requirements. The method is applicable to various scenarios, including resource-efficient training and rendering, quality-prioritized rendering, and storage compression. The authors also discuss limitations and future work, including the potential for further refinement through the integration of stereo matching techniques.Mini-Splatting: Representing Scenes with a Constrained Number of Gaussians This paper introduces Mini-Splatting, a method for efficiently representing scenes using a constrained number of Gaussians. The key challenge is to maintain high-quality rendering while reducing the number of Gaussians. The authors analyze the inefficiencies of Gaussian representation in 3D Gaussian Splatting (3DGS), where Gaussians are often clustered, leading to suboptimal rendering quality and speed. To address this, they propose a densification and simplification algorithm that reorganizes the spatial positions of Gaussians, resulting in a more uniform distribution and improved performance. The densification process includes blur split and depth reinitialization. Blur split involves splitting Gaussians in areas with smooth color transitions, while depth reinitialization uses merged depth points to reinitialize the scene. The simplification process includes intersection preserving, which retains Gaussians with the most impact on the rendered image, and sampling, which preserves both geometric structure and rendering quality. The proposed methods are integrated into the original rasterization pipeline, providing a strong baseline for future research in Gaussian-Splatting-based works. The results show that Mini-Splatting achieves a balanced trade-off between rendering quality, resource consumption, and storage. The method is evaluated on multiple benchmarks and datasets, demonstrating its potential and scalability. The authors also discuss related work, including Gaussian Splatting and 3D data simplification. They highlight the challenges of Gaussian simplification, such as inefficient spatial distribution and the need for effective pruning strategies. The paper presents ablation studies and quantitative results, showing that their methods outperform existing approaches in terms of rendering quality and efficiency. The paper concludes that Mini-Splatting effectively represents scenes with a constrained number of Gaussians, achieving high-quality rendering while reducing computational and storage requirements. The method is applicable to various scenarios, including resource-efficient training and rendering, quality-prioritized rendering, and storage compression. The authors also discuss limitations and future work, including the potential for further refinement through the integration of stereo matching techniques.