Research map / Label-efficient and robust learning / Semi-supervised learning
PanDA: Towards Panoramic Depth Anything with Unlabeled Panoramas and Mobius Spatial Augmentation
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.
Analyzes Depth Anything Models on panoramas, then builds PanDA, a semi-supervised panoramic depth model (LoRA-tuned DAM v2 teacher, student trained on unlabeled panoramas) with Möbius-transformation spatial consistency augmentation.
Problem
Depth Anything Models perform well on perspective images but their behavior on 360° panoramas (spherical distortion, camera position, spherical transformations) is unclear and they are sensitive to spatial transformations; panoramic depth labels are scarce and scene-specific.
Approach
Empirically analyze DAMs on panoramas, then train a teacher via LoRA fine-tuning on synthetic indoor/outdoor panoramas with an equator-aware patch normalization loss, and a student on large unlabeled panoramas with Möbius-transformation-based spatial augmentation enforcing depth consistency.
Key results
- PanDA-L reaches AbsRel 0.0717 and RMSE 0.3305 on Matterport3D, outperforming HRDFuse (RMSE 0.4433) and other panoramic methods.
- On Stanford2D3D, PanDA-L achieves AbsRel 0.0609 and RMSE 0.2540, improving over HRDFuse (AbsRel 0.0935, RMSE 0.3106).
- In zero-shot evaluation with vertical rotation of 20° and zoom 3.0, PanDA shows much smaller error degradation than DAM v2 (e.g., RMSE 0.7461 vs 1.0433 on Matterport3D ViT-S setting).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune DAM v2 and train with panoramic data | DAM v2 · Matterport3D | AbsRel: 0.0850 -> 0.0717 | Helped |
| fine-tune DAM v2 and train with panoramic data | DAM v2 · Stanford2D3D | RMSE: 0.3106 -> 0.2540 | Helped |
| fine-tune DAM v2 for panoramic depth estimation | DAM v2 ViT-L · Matterport3D, original | RMSE: 0.5522 -> 0.4690 | Helped |
| add EPNL (ablation) | PanDA · Matterport3D | RMSE: 0.5132 -> 0.5062 | Helped |
| add EPNL (ablation) | PanDA · Stanford2D3D | RMSE: 0.3418 -> 0.3401 | Helped |
| add EPNL (ablation) | PanDA · Matterport3D | AbsRel: 0.1057 -> 0.1036 | Helped |
| add EPNL (ablation) | PanDA · Stanford2D3D | AbsRel: 0.1082 -> 0.1092 | Hurt |
| replace EPNL with RPNL (ablation) | PanDA · Matterport3D | RMSE: 0.5132 -> 0.5268 | Hurt |
| replace EPNL with RPNL (ablation) | PanDA · Stanford2D3D | RMSE: 0.3418 -> 0.3533 | Hurt |
| train using pseudo-labels from unlabeled panoramas (ablation) | PanDA student · Matterport3D, original | RMSE: 0.5109 -> 0.4977 | Helped |
| add MTSA consistency loss (ablation) | PanDA student · Matterport3D, original | RMSE: 0.4977 -> 0.4950 | Helped |
| add color augmentation consistency loss without MTSA (ablation) | PanDA student · Matterport3D, original | RMSE: 0.4977 -> 0.4984 | No clear effect |
Surprising findings
- Higher zoom levels above about 2.4 can improve DAM performance on panoramas, attributed to enlarged regions simplifying scene structure.
- Using ERP as input is best among tested panoramic representations for DAMs, even though it has the most distortion.
Limitations
- Performance degrades with large vertical rotations and polar-dominated panoramas, per the analysis.
- Fine-tuning on real-world datasets is needed to match specialized methods in some settings; zero-shot advantage is measured on limited indoor benchmarks.
Experimental setup
Datasets: Structured3D Deep360 ZInD 360+x Matterport3D Stanford2D3D
Benchmarks: Matterport3D Stanford2D3D
Metrics: AbsRel RMSE delta1 delta2 delta3
Baselines: Depth Anything v1 Depth Anything v2 Marigold Depth Anywhere UniFuse HRDFuse S2Net ACDNet PanoFormer EGFormer Elite360D BiFuse
Builds on: Depth Anything v2 Depth Anything v1 LoRA Depth Anywhere UniFuse
Closest papers by meaning
- DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images (NeurIPS 2026 · similarity 0.88)
- Unified Panoramic Geometry Estimation via Multi-View Foundation Models (NeurIPS 2026 · similarity 0.87)
- Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data (CVPR 2024 · similarity 0.86)
- Splatter-360: Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images (CVPR 2025 · similarity 0.86)
- RSA: Resolving Scale Ambiguities in Monocular Depth Estimators through Language Descriptions (NeurIPS 2024 · similarity 0.85)
- Depth Anything V2 (NeurIPS 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.