Research map / 3D vision / Gaussian splatting
ActiveGAMER: Active GAussian Mapping through Efficient Rendering
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 ActiveGAMER, a 3DGS-based active mapping system that uses rendering-based information gain for next-best-view planning, with coarse-to-fine exploration, post-refinement, and global-local keyframe selection.
Problem
Active reconstruction (next-best-view exploration) with radiance fields is limited by NeRF's slow rendering, and prior work neglects photometric quality while focusing on geometry; the goal is complete, accurate, and photorealistic scene mapping with unrestricted 6DoF motion.
Approach
Uses the fast rendering of a 3D Gaussian map (built on SplaTAM) to compute silhouette-based information gain over sampled candidate viewpoints, combined with a coarse-to-fine candidate sampling schedule, a post-refinement stage for rendering quality, and a global-local keyframe selection strategy to counter local overfitting.
Key results
- On MP3D, ActiveGAMER reaches 1.66 cm accuracy, 2.30 cm completion, and 95.32% completion ratio, versus 6.31 cm / 3.00 cm / 90.18% for NARUTO.
- On Replica novel view rendering, ActiveGAMER achieves 32.02 dB average PSNR versus 29.08 dB for SplaTAM and 26.01 dB for NARUTO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add rendering-based active mapping with 3D Gaussian Splatting | ActiveGAMER · MP3D | Completion ratio (%): 90.18 -> 95.32 | Helped |
| add rendering-based active mapping with 3D Gaussian Splatting | ActiveGAMER · Replica (Avg.) | PSNR: 29.08 -> 32.02 | Helped |
| add fine exploration after coarse exploration (ablation) | ActiveGAMER · Replica | PSNR (dB): 29.77 -> 32.02 | Helped |
| add fine exploration after coarse exploration (ablation) | ActiveGAMER · Replica | Completion (cm): 1.77 -> 1.80 | Mixed |
| add global keyframes to local keyframe selection (ablation) | ActiveGAMER · Replica | PSNR (dB): 30.73 -> 32.02 | Helped |
| add global keyframes to local keyframe selection (ablation) | ActiveGAMER · Replica | Completion (cm): 2.19 -> 1.80 | Helped |
| add post-refinement (ablation) | ActiveGAMER · Replica | PSNR (dB): 30.67 -> 32.02 | Helped |
| add post-refinement (ablation) | ActiveGAMER · Replica | Completion ratio (%): 96.50 -> 95.45 | Hurt |
Surprising findings
- Post-refinement improves rendering quality but reduces geometric completeness due to Gaussian pruning, so the paper uses separate exploration and refinement models for geometry and rendering evaluation.
Limitations
- Assumes known localization and perfect action execution; not validated in real-world settings.
- Rendering-based information gain may miss regions such as back sides of objects not visible from initial views.
Experimental setup
Datasets: Replica Matterport3D
Benchmarks: Replica MP3D
Metrics: Accuracy (cm) Completion (cm) Completion ratio (%) PSNR SSIM LPIPS Depth L1
Baselines: FBE UPEN OccAnt ANM NARUTO SplaTAM
Builds on: SplaTAM NARUTO 3D Gaussian Splatting RRT
Closest papers by meaning
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.88)
- Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.88)
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.88)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.87)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · 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.