Research map / 3D vision / 3D generation
Doppelgangers++: Improved Visual Disambiguation with Geometric 3D Features
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.
Doppelgangers++ trains a Transformer classifier on frozen MASt3R decoder features plus a new VisymScenes dataset to detect visually aliased image pairs and prune them before SfM.
Problem
Visual aliasing (doppelgangers): visually similar but distinct surfaces produce spurious feature matches that corrupt structure-from-motion reconstructions; prior CNN classifiers generalize poorly beyond landmark photos and are sensitive to thresholds.
Approach
Reuse multi-layer, 3D-aware decoder features from a frozen MASt3R model, feed them into two Transformer classification heads on symmetrized image pairs with test-time voting, and expand training with geo-tagged everyday-scene data (VisymScenes).
Key results
- On the out-of-domain VisymScenes test set, the method reaches AP 0.991 and ROC AUC 0.990 when trained on DG plus VisymScenes, versus 0.938 and 0.921 for DG-OG.
- On Mapillary (out-of-domain), trained on DG plus VisymScenes, AP is 0.968 versus 0.692 for DG-OG.
- On SfM reconstruction, the method succeeds on Belvedere (Vienna) where DG-OG fails (inlier ratio 0.874 vs 0.451), using a consistent threshold of 0.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Doppelgangers++ instead of DG-OG, trained on DG | pairwise visual disambiguation · Mapillary test | AP: 0.566 -> 0.950 | Helped |
| Use Doppelgangers++ instead of DG-OG, trained on DG | pairwise visual disambiguation · VisymScenes test | Prec@Recall=0.85: 0.498 -> 0.892 | Helped |
| Add VisymScenes training pairs | DG-OG · Mapillary test | AP: 0.566 / 0.692 | Helped |
| Add VisymScenes training pairs | Doppelgangers++ · Mapillary test | AP: 0.950 / 0.968 | Helped |
| Use Doppelgangers++ instead of DG-OG | COLMAP SfM reconstruction · Belvedere (Vienna) | Inlier Ratio: 0.451 -> 0.874 | Helped |
| Use Doppelgangers++ instead of COLMAP | SfM reconstruction · Belvedere (Vienna) | Inlier Ratio: 0.521 -> 0.874 | Helped |
| Fine-tune the entire MASt3R model instead of training only the classification heads (ablation) | Doppelgangers++ · pairwise visual disambiguation test sets | Recall | Hurt |
| Use one classification head on combined decoder-branch features instead of two separate heads (ablation) | Doppelgangers++ · pairwise visual disambiguation test sets | classification performance | Hurt |
| Replace the Transformer classification head with an MLP (ablation) | Doppelgangers++ · pairwise visual disambiguation test sets | classification performance | Hurt |
| Use only final-layer decoder features instead of multi-layer decoder features (ablation) | Doppelgangers++ · pairwise visual disambiguation test sets | classification performance | Hurt |
Surprising findings
- Adding VisymScenes training data does not improve DG-OG on the out-of-domain Mapillary test set, while the proposed method keeps benefiting.
- Fine-tuning the full MASt3R model led to lower recall, suggesting overfitting, so training only the heads was preferred.
Limitations
- Precision outside the training domain was suboptimal when trained only on DG, requiring the expanded dataset.
- Geo-alignment inlier ratio depends on the accuracy of geotags, so it is only comparable within the same scene.
- Requires per-scene or per-dataset threshold choices in baselines; the method still uses a threshold tau of 0.8.
Experimental setup
Datasets: Doppelganger dataset (DG) VisymScenes Mapillary Street-Level Sequences MegaScenes
Benchmarks: DG test set VisymScenes test set Mapillary test set SfM scene reconstruction on 21 scenes
Metrics: Average Precision ROC AUC precision at fixed recall recall at fixed precision registered image count geo-alignment inlier ratio
Baselines: DG-OG (Doppelgangers) COLMAP MASt3R-SfM SIFT+RANSAC SuperPoint+SuperGlue
Builds on: MASt3R Doppelgangers COLMAP MASt3R-SfM
Closest papers by meaning
- Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation (NeurIPS 2025 · similarity 0.86)
- RefineAny3D: Depth Refinement as Semantic Alignment for Monocular 3D Detection (NeurIPS 2026 · similarity 0.85)
- Tango3D: Towards Alignment for Global and Local 2D-3D Correspondence (NeurIPS 2026 · similarity 0.85)
- SIU3R: Simultaneous Scene Understanding and 3D Reconstruction Beyond Feature Alignment (NeurIPS 2025 · similarity 0.85)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.85)
- SPARS3R: Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction (CVPR 2025 · 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.