Research map / 3D vision / Depth and stereo
SCRREAM : SCan, Register, REnder And Map: A Framework for Annotating Accurate and Dense 3D Indoor Scenes with a Benchmark
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 SCRREAM, a pipeline that pre-scans objects into watertight meshes, registers them into a scene, and maps a free-hand RGB(-D) sequence to synthetic renderings to get camera poses, producing dense ground-truth depth and a benchmark for NVS and SLAM.
Problem
Indoor 3D datasets prioritize scale over ground-truth accuracy; their incomplete meshes yield wrong depth ground truth for dense geometry evaluation such as depth rendering.
Approach
Combine individually pre-scanned high-resolution object meshes, registered into a partially rescanned scene via manual correspondences and ICP, with feature-matching-based camera pose estimation (two-stage COLMAP-style mapping onto synthetic renderings) instead of external tracking, enabling dense, complete ground truth with free-hand camera trajectories.
Key results
- On NVS, Depth-Nerfacto with active stereo depth prior achieved Abs Rel 0.079 and delta<1.25 0.968, versus 0.645 Abs Rel for NeRFacto without depth prior.
- On SLAM with ground-truth depth, CO-SLAM reached ATE 4.18 cm and Completion Ratio 96.52%, while with ToF depth its ATE rose to 45.94 cm.
- Zip-NeRF achieved the best RGB PSNR of 28.315 but a depth Abs Rel of 0.245, showing photometric quality does not track depth quality.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add an Active Stereo depth prior to NeRFacto | NeRFacto / Depth-Facto · SCRREAM indoor scenes | RMSE (depth): 1.244 -> 0.218 | Helped |
| Add a ToF depth prior to NeRFacto | NeRFacto / Depth-Facto · SCRREAM indoor scenes | RMSE (depth): 1.244 -> 0.336 | Helped |
| Use ground-truth depth input instead of Active Stereo depth | NICE-SLAM · SCRREAM indoor scenes | ATE [cm]: 39.09 -> 5.47 | Helped |
| Use ground-truth depth input instead of ToF depth | NICE-SLAM · SCRREAM indoor scenes | ATE [cm]: 24.02 -> 5.47 | Helped |
| Use ground-truth depth input instead of Active Stereo depth | CO-SLAM · SCRREAM indoor scenes | ATE [cm]: 39.19 -> 4.18 | Helped |
| Use ground-truth depth input instead of ToF depth | CO-SLAM · SCRREAM indoor scenes | ATE [cm]: 45.94 -> 4.18 | Helped |
| Use ground-truth depth input instead of Active Stereo depth | Gaussian-SLAM · SCRREAM indoor scenes | ATE [cm]: 38.20 -> 5.00 | Helped |
| Use ground-truth depth input instead of ToF depth | Gaussian-SLAM · SCRREAM indoor scenes | ATE [cm]: 83.53 -> 5.00 | Helped |
| Use ToF depth instead of Active Stereo depth | NICE-SLAM · SCRREAM indoor scenes | ATE [cm]: 39.09 -> 24.02 | Helped |
| Use ToF depth instead of Active Stereo depth | CO-SLAM · SCRREAM indoor scenes | ATE [cm]: 39.19 -> 45.94 | Hurt |
| Use ToF depth instead of Active Stereo depth | Gaussian-SLAM · SCRREAM indoor scenes | ATE [cm]: 38.20 -> 83.53 | Hurt |
| Use sensor depth to supervise the NVS method | Depth-Facto · SCRREAM indoor scenes | depth rendering quality | Helped |
Surprising findings
- Zip-NeRF has the highest PSNR (28.315) yet much worse depth accuracy than depth-supervised NeRFacto variants, so photometric quality does not coincide with depth quality.
- With sensor depth (ToF) as input, SLAM tracking errors rise sharply (e.g., CO-SLAM ATE from 4.18 cm with GT depth to 45.94 cm with ToF).
Limitations
- Hardware setup is expensive, complex, and time-consuming, making the pipeline less scalable.
- Fixed exposure and static capture requirements cause under/over-exposure, motion blur, and limit scene edits to object removal without human interaction.
- Human reconstruction uses a stop-motion mannequin instead of real humans, and the human and 6D pose variants lack benchmarks.
Experimental setup
Datasets: SCRREAM (eleven indoor scenes)
Benchmarks: SCRREAM indoor reconstruction and SLAM benchmark
Metrics: PSNR SSIM LPIPS RMSE Abs Rel Sqr Rel delta<1.25 ATE Accuracy Completion
Baselines: NeRFacto Depth-Nerfacto Zip-NeRF 3D Gaussian Splatting Mip-Splatting NICE-SLAM CO-SLAM Gaussian-SLAM
Builds on: COLMAP Horn's method ICP HAMMER HouseCat6D
Closest papers by meaning
- Rooms from Motion: Un-posed Indoor 3D Object Detection as Localization and Mapping (NeurIPS 2025 · similarity 0.86)
- HouseLayout3D: A Benchmark and Training-free Baseline for 3D Layout Estimation in the Wild (NeurIPS 2025 · similarity 0.86)
- PLANA3R: Zero-shot Metric Planar 3D Reconstruction via Feed-forward Planar Splatting (NeurIPS 2025 · similarity 0.85)
- InternScenes: A Large-scale Simulatable Indoor Scene Dataset with Realistic Layouts (NeurIPS 2025 · similarity 0.85)
- PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes (CVPR 2025 · similarity 0.85)
- LASA: Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset (CVPR 2024 · similarity 0.85)
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.