19 August 2026 in SSRN Electronic Journal
Large-scale 3D Gaussian reconstruction is constrained by GPU memory and by the difficulty of representing spatially heterogeneous scenes with one model. Partitioned training alleviates these constraints, but fixed data-assignment rules do not transfer cleanly between aerial and forward-facing captures, and independently trained Scaffold-GS partitions cannot share a decoder without corrupting anchor semantics. We present BlockScaffold, a partitioned Scaffold-GS framework that uses camera-pose statistics to select scene-dependent camera assignment and boundary handling, chooses an aspect-ratio-aware grid for a user-specified partition budget, and routes every anchor through the decoder of its source partition. The routing rule follows from the non-identifiability of independently learned anchor feature spaces and preserves the anchor--decoder pairing established during training. Across six scenes from Mill19, UrbanScene3D, and MatrixCity, BlockScaffold achieves the highest evaluated PSNR on three of five aerial scenes and the highest SSIM on two, including both metrics on MatrixCity-Aerial. On Rubble, partitioned training reduces peak memory from approximately 22~GB for Scaffold-GS to 6.6~GB per partition while rendering at 57.1~frames/s. Source-aware routing improves MatrixCity-Aerial PSNR by more than 6~dB over single-decoder merging, demonstrating that decoder compatibility is critical when neural-anchor partitions are combined. On the evaluated street scene, the method improves over Scaffold-GS while leaving scope for stronger cross-partition appearance alignment.
29 September 2026 in SSRN Electronic Journal
GOGabriel Leivas Oliveira
Large-scale 3D Gaussian reconstruction is constrained by GPU memory and by the difficulty of representing spatially heterogeneous scenes with one model. Partitioned training alleviates these constraints, but fixed data-assignment rules do not transfer cleanly between aerial and forward-facing captures, and independently trained Scaffold-GS partitions cannot share a decoder without corrupting anchor semantics. We present BlockScaffold, a partitioned Scaffold-GS framework that uses camera-pose statistics to select scene-dependent camera assignment and boundary handling, chooses an aspect-ratio-aware grid for a user-specified partition budget, and routes every anchor through the decoder of its source partition. The routing rule follows from the non-identifiability of independently learned anchor feature spaces and preserves the anchor--decoder pairing established during training. Across six scenes from Mill19, UrbanScene3D, and MatrixCity, BlockScaffold achieves the highest evaluated PSNR on three of five aerial scenes and the highest SSIM on two, including both metrics on MatrixCity-Aerial. On Rubble, partitioned training reduces peak memory from approximately 22~GB for Scaffold-GS to 6.6~GB per partition while rendering at 57.1~frames/s. Source-aware routing improves MatrixCity-Aerial PSNR by more than 6~dB over single-decoder merging, demonstrating that decoder compatibility is critical when neural-anchor partitions are combined. On the evaluated street scene, the method improves over Scaffold-GS while leaving scope for stronger cross-partition appearance alignment.
23 June 2026 in SSRN Electronic Journal
Recent advances in 3D scene reconstruction have highlighted the intricate trade-offs among rendering quality, inference efficiency, and data dependency. To address the challenge of rapidly reconstructing detailed 3D indoor scenes from minimal input, we introduce FastPano3D, an end-to-end framework that directly generates renderable 3D Gaussian representations from a single panoramic image. Unlike perspective-based methods, panoramic images inherently suffer from equirectangular projection distortions and spatially non-uniform feature distributions, making direct feed-forward Gaussian generation particularly challenging. In contrast to existing Gaussian Splatting based methods that rely on multi-view supervision or per-scene optimization, FastPano3D employs a lightweight feature encoder, adaptive Gaussian sampling, and a point-cloud-guided refinement strategy to achieve efficient and accurate scene generation without any test-time optimization. Our approach reconstructs high-fidelity 3D scenes within seconds, achieving up to 156× faster inference than prior state-of-the-art methods such as Pano2Room, while using only half the parameters. Extensive experiments demonstrate that FastPano3D delivers rendering quality comparable to NeRF- and 3DGS-based reconstructions, establishing a new benchmark for rapid, single-view 3D scene inference.
25 September 2026 in SSRN Electronic Journal
3D Gaussian Splatting (3DGS) enables high-quality real-time novel-view synthesis but incurs substantial storage overhead due to the large number of Gaussian primitives required for complex scenes. Existing compression methods either rely on expensive per-scene retraining or apply uniform post-training strategies that ignore the heterogeneous importance of individual Gaussians. We present FA-Compress, a training-free post-processing framework that exploits this heterogeneity: primitives in smooth regions tolerate substantially more compression than those encoding high-frequency structures. FA-Compress estimates per-Gaussian frequency characteristics directly from model-internal attributes and performs frequency-guided SH truncation, opacity pruning, and adaptive quantization without retraining or rendering-based optimization. On NeRF Synthetic, Mip-NeRF 360, and Tanks & Temples, FA-Compress consistently achieves higher rendering quality than existing training-free methods and competitive quality against retraining-based approaches, completing in seconds while remaining fully compatible with the standard 3DGS rendering pipeline.