Research map / 3D vision / Gaussian splatting
GaussianBlock: Building Part-Aware Compositional and Editable 3D Scene by Primitives and Gaussians
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GaussianBlock, a hybrid superquadric-plus-3D-Gaussian representation for part-aware, editable compositional 3D reconstruction from multi-view images, guided by SAM attention clustering.
Problem
Existing 3D reconstruction representations are entangled; primitive-based methods lack fidelity and semantic coherence, while Gaussian-based segmentation methods only support holistic edits rather than part-level controllable editing.
Approach
Combine superquadric primitives (coarse, disentangled editable blocks) with Gaussians bound to primitive triangles via binding inheritance; supervise superquadric semantic coherence with an attention-guided centering loss using SAM cross-attention clustered by HDBSCAN, plus dynamic splitting and fusion.
Key results
- On DTU primitive reconstruction, mean Chamfer Distance 3.69 vs 4.78 for DBW and 5.67 for EMS (EMS uses GT point clouds).
- Full model reaches 30.7 PSNR / 94.3 SSIM on DTU, compared with 35.7 PSNR / 98.3 SSIM for 3DGS, so fidelity is competitive but not best.
- Ablation shows removing the AC loss raises CD from 3.69 to 4.75 and removing splitting drops PSNR from 30.7 to 25.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GaussianBlock instead of DBW primitives | DTU primitive reconstruction · DTU (S24–S105) | CD: 4.78 -> 3.69 | Helped |
| Use GaussianBlock instead of 3DGS | GaussianBlock · DTU | SSIM: 98.3 -> 94.3 | Hurt |
| Use GaussianBlock instead of 3DGS | GaussianBlock · BlendedMVS | PSNR: 33.8 -> 28.8 | Hurt |
| Use GaussianBlock instead of 3DGS | GaussianBlock · Truck | LPIPS: 17.3 -> 20.7 | Hurt |
| Use GaussianBlock instead of 3DGS | GaussianBlock · Garden | SSIM: 86.2 -> 84.8 | Hurt |
| Add attention-guided centering loss (ablation) | GaussianBlock · DTU | CD: 4.75 -> 3.69 | Helped |
| Add dynamic splitting (ablation) | GaussianBlock · DTU | CD: 4.24 -> 3.69 | Helped |
| Add dynamic fusion (ablation) | GaussianBlock · DTU | CD: 3.69 -> 3.46 | Hurt |
| Set AC-loss weight γ to 0.02 instead of 0.2 (ablation) | GaussianBlock · DTU | CD: 3.69 -> 4.12 | Hurt |
| Set AC-loss weight γ to 2 instead of 0.2 (ablation) | GaussianBlock · DTU | CD: 3.69 -> 6.97 | Hurt |
| Set splitting threshold β to 0.5 instead of 0.7 (ablation) | GaussianBlock · DTU | CD: 3.69 -> 3.56 | Helped |
| Set splitting threshold β to 0.9 instead of 0.7 (ablation) | GaussianBlock · DTU | CD: 3.69 -> 4.85 | Hurt |
Surprising findings
- Fidelity is noticeably lower than pure Gaussian baselines (about 5 PSNR below 3DGS on DTU), which the authors accept as a trade-off for part-level editability.
Limitations
- Only object-centric entities within the silhouette can be decomposed; backgrounds are not handled well.
- Superquadric representation is coarse, so thin or detailed structures such as leaves and wood fail.
- Fidelity is below purely Gaussian-based methods.
Experimental setup
Datasets: DTU BlendedMVS Nerfstudio Mip-360 Tanks and Temples
Benchmarks: DTU BlendedMVS Tanks and Temples Truck Mip-360 Garden
Metrics: Chamfer Distance SSIM PSNR LPIPS
Baselines: 3DGS OmniSeg3D DBW EMS
Builds on: 3D Gaussian Splatting GaussianAvatar SAM HDBSCAN DBW
Closest papers by meaning
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.89)
- Learning Efficient Fuse-and-Refine for Feed-Forward 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.88)
- 3DitScene: Editing Any Scene via Language-guided Disentangled Gaussian Splatting (ICLR 2025 · similarity 0.88)
- CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction (NeurIPS 2026 · similarity 0.88)
- Tackling View-Dependent Semantics in 3D Language Gaussian Splatting (ICML 2025 · similarity 0.88)
- LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.