Research map / 3D vision / Gaussian splatting
Multi-View Pose-Agnostic Change Localization with Zero Labels
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 a label-free, pose-agnostic multi-view change detection method that embeds feature- and SSIM-based change masks as extra change channels in a 3D Gaussian Splatting model, and releases the PASLCD real-world dataset.
Problem
Detecting and localizing scene changes from unconstrained, inconsistent viewpoints without labeled training data or pose alignment between pre- and post-change observations.
Approach
Build a 3DGS of the reference scene, compare rendered and observed images using DINOv2 feature differences combined with SSIM to form per-view change masks, then learn two extra per-Gaussian change channels (degree-0 SH change color and change opacity) in a post-change 3DGS so change masks are consistent across views and renderable for unseen poses.
Key results
- On ChangeSim the method reaches 0.407 mIoU on the Change class vs 0.229 for CSCDNet (about 1.7x).
- On PASLCD, the method reaches 0.461 mIoU and 0.612 F1 averaged over scenes, outperforming SplatPose (0.173 mIoU) and CYWS-2D (0.273 mIoU in pair-wise setting).
- With only 5 inference images, the method achieves roughly 1.8x the mIoU of the Feature Diff baseline, and unseen-view masks reach 0.36-0.45 mIoU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed multi-view change-aware 3DGS method instead of SplatPose | MAD-Real pose-agnostic change localization · MAD-Real | mIoU: 0.077 -> 0.132 | Helped |
| Use the proposed multi-view change-aware 3DGS method instead of SplatPose | MAD-Real pose-agnostic change localization · MAD-Real | F1: 0.123 -> 0.210 | Helped |
| Use the proposed method instead of Feature Difference | MAD-Real pose-agnostic change localization · MAD-Real | AUROC: 0.967 -> 0.953 | Hurt |
| Use the proposed multi-view method instead of Feature Difference | ChangeSim change localization · ChangeSim Change class | mIoU: 0.352 -> 0.407 | Helped |
| Use the proposed method instead of CYWS-2D | PASLCD scene-level change localization · PASLCD | mIoU: 0.273 -> 0.461 | Helped |
| Use the proposed method instead of CYWS-2D | PASLCD scene-level change localization · PASLCD | F1: 0.398 -> 0.612 | Helped |
| Add Change-3DGS multi-view consistency to CYWS-2D | CYWS-2D change masks · PASLCD | mIoU: 0.273 -> 0.393 | Helped |
| Use degree 0 rather than degree 3 spherical harmonics for change (ablation) | Change-3DGS · PASLCD indoor scenes | mIoU: 0.442 -> 0.4741 | Helped |
| Add Change-3DGS learning from feature-aware masks (ablation) | Feature Difference · PASLCD | mIoU: 0.264 -> 0.311 | Helped |
| Add Change-3DGS learning from structure-aware masks (ablation) | Feature Difference · PASLCD | mIoU: 0.264 -> 0.324 | Helped |
| Combine feature-aware and structure-aware masks instead of using either alone (ablation) | Change-3DGS · PASLCD | mIoU: 0.311 -> 0.449; 0.324 -> 0.449 | Helped |
| Add reference-pose data augmentation (ablation) | Combined Change-3DGS · PASLCD | mIoU: 0.449 -> 0.457 | Helped |
Surprising findings
- Lower spherical harmonics degree for change channels (degree 0) reduced false positives by about 70% and gave the best mIoU, contrary to the default higher-degree choice.
Limitations
- Fails to detect color-based or surface-level changes because the DINOv2 feature masks miss them.
- Struggles with very small changes in large-scale scenes.
- Feature masks overestimate true change regions due to patch-to-pixel interpolation; convergence of baselines depends on pose overlap.
Experimental setup
Datasets: PASLCD ChangeSim MAD-Real
Benchmarks: PASLCD ChangeSim MAD-Real
Metrics: mIoU F1 AUROC
Baselines: OmniPoseAD SplatPose Feature Difference CYWS-2D CYWS-3D CSCDNet ChangeNet
Builds on: 3D Gaussian Splatting DINOv2 SplatPose OmniPoseAD CYWS-2D
Closest papers by meaning
- Living Scenes: Multi-object Relocalization and Reconstruction in Changing 3D Environments (CVPR 2024 · similarity 0.87)
- Towards Generalizable Scene Change Detection (CVPR 2025 · similarity 0.87)
- HUGS: Holistic Urban 3D Scene Understanding via Gaussian Splatting (CVPR 2024 · similarity 0.86)
- EGSRAL:An Enhanced 3D Gaussian Splatting Based Renderer with Automated Labeling for Large-Scale Driving Scene (AAAI 2025 · similarity 0.86)
- FreeSplat: Generalizable 3D Gaussian Splatting Towards Free View Synthesis of Indoor Scenes (NeurIPS 2024 · similarity 0.86)
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.86)
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.