Multi-Scale 3D Gaussian Splatting with Region-Adaptive Structure Loss and Scale-Wise Weighting = Region-Adaptive 구조 손실과 스케일 적응형 가중을 적용한 다중 스케일 가우시안 스플래팅

다국어 초록 (Multilingual Abstract)

By employing anisotropic Gaussian primitives, 3D Gaussian splatting （3DGS） achieves real-time novel view synthesis. A common limitation is that rendering quality tends to deteriorate once a single trained representation is displayed across several scales. Prior multi-scale 3DGS approaches concentrate chiefly on adjusting the rendering operator to suppress aliasing, yet their training objectives continue to depend on uniform pixel-wise supervision together with fixed scale weights. Under such conditions, structurally intricate regions and challenging rendering scales may receive insufficient emphasis. We tackle this imbalance on the training side by introducing the notion of regional-scale image-space complexity （RSIC） and, building on this notion, we present RSIC-aware 3D Gaussian splatting （RSIC-GS） as a strategy for training. This strategy couples a region-adaptive structure loss （RASL） with a scale-adaptive weighting （SAW） scheme: RASL relies on a Laplacian-derived RSIC measure that highlights structurally intricate regions, whereas SAW assigns a scale-specific weight to every rendering scale within the training loss. Our ablation experiments indicate that RASL and SAW are best optimized together, since weighting across scales lets the structural loss enhance recovery at coarse scales while curbing its penalty at fine scales. We integrate the proposed method into MS-GS as well as Mip-Splatting without altering how rendering proceeds at inference time. The experiments demonstrate that, for the retrained MS-GS backbone, RSIC-GS lifts the four-scale mean PSNR by 0.51 dB on Mip-NeRF 360, by 0.36 dB on Tanks and Temples, and by 0.14 dB on Deep Blending, while keeping per-frame runtime comparable and incurring negligible storage overhead.

목차 (Table of Contents)

• Ⅰ. Introduction 1
• Ⅱ. Related Works 7
• 2.1 NeRF-based Multi-Scale Methods 9
• 2.2 3DGS-based Multi-Scale Methods 11
• 2.3 Loss Functions for Multi-Scale Rendering 13
• Ⅲ. Proposed Methods 19
• 3.1 Overall Framework 19
• 3.2 Region-Adaptive Structure Loss （RASL） 24
• 3.3 Scale-Adaptive Weighting （SAW） 26
• Ⅳ. Experimental Results 29
• 4.1 Experimental Setup 29
• 4.2 Multi-Scale Rendering Performance 32
• 4.3 Ablation Studies 41
• Ⅴ. Conclusion 48
• Reference 50
• 국문초록 59
• Acknowledgement （in Korean） 61
• Curriculum Vitae （in Korean） 63