Research map / 3D vision / Depth and stereo
SO(3)-Equivariant ViT-Adapter for Data-Efficient Zero-Shot Sim-to-Real Indoor Panoramic Depth Estimation
Ziyan He, Qiudan Zhang, Lin Ma, Xu Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace perspective-only or semi-supervised panoramic baselines with SO(3)-equivariant ViT-Adapter trained on synthetic data only | Depth Anything V2 ViT-L · Matterport3D (zero-shot sim-to-real) | Abs Rel: 0.091 (PanDA-L) -> 0.078 | Improved |
| replace perspective-only or semi-supervised panoramic baselines with SO(3)-equivariant ViT-Adapter | Depth Anything V2 ViT-L · Stanford2D3D (zero-shot sim-to-real) | Abs Rel: 0.073 (PanDA-L) -> 0.060 | Improved |
| add a plain ViT-Adapter instead of full fine-tuning of the ViT-S encoder (ablation) | Depth Anything V2 ViT-S · Matterport3D (zero-shot sim-to-real) | RMSE: 0.439 -> 0.431 | Improved |
| fine-tune the whole encoder instead of keeping it frozen (ablation) | Depth Anything V2 ViT-S · Matterport3D (zero-shot sim-to-real) | Abs Rel: 0.100 -> 0.093 | Improved |
| add the ESPM equivariant prior without SO(3)-DCAM (ablation) | ViT-Adapter · Matterport3D (zero-shot sim-to-real) | Abs Rel: 0.093 -> 0.092 | Improved |
| add SO(3)-DCAM on top of ESPM (ablation) | SO(3)-Equivariant ViT-Adapter (w/o SCRPE) · Matterport3D (zero-shot sim-to-real) | Abs Rel: 0.092 -> 0.091 (w/o SCRPE) -> 0.089 (full) | Improved |
| add SCRPE relative positional encoding (ablation) | SO(3)-Equivariant ViT-Adapter · Matterport3D (zero-shot sim-to-real) | RMSE: 0.426 -> 0.416 | Improved |
| use full SO(3)-equivariant adapter instead of plain ViT-Adapter at 20 deg vertical rotation, no rotation augmentation (ablation) | ViT-S · Matterport3D, vertical rotation 20 deg | RMSE: 0.690 -> 0.612 | Improved |
| add vertical rotation augmentation during training (ablation) | ViT-Adapter · Matterport3D, vertical rotation 20 deg | RMSE: 0.690 -> 0.593 (14.06%) | Improved |
| add vertical rotation augmentation during training (ablation) | SO(3)-Equivariant ViT-Adapter · Matterport3D, vertical rotation 20 deg | RMSE: 0.612 -> 0.556 (9.15%) | Improved |
| add vertical rotation augmentation during training (ablation) | SO(3)-Equivariant ViT-Adapter · Matterport3D, no rotation (theta = 0) | RMSE: 0.416 -> 0.425 (-2.16%) | Hurt |
| add vertical rotation augmentation during training (ablation) | SO(3)-Equivariant ViT-Adapter · Matterport3D, no rotation (theta = 0) | delta1: 0.882 -> 0.874 (-0.91%) | Hurt |
Closest papers by meaning
- VGGT-360: Geometry-Consistent Zero-Shot Panoramic Depth Estimation (CVPR 2026 · similarity 0.92)
- Unified Panoramic Geometry Estimation via Multi-View Foundation Models (NeurIPS 2026 · similarity 0.88)
- DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images (NeurIPS 2026 · similarity 0.88)
- Splatter-360: Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images (CVPR 2025 · similarity 0.87)
- CM-EVS: Sparse Panoramic RGB-D-Pose Data for Complete Scene Coverage (NeurIPS 2026 · similarity 0.87)
- PanDA: Towards Panoramic Depth Anything with Unlabeled Panoramas and Mobius Spatial Augmentation (CVPR 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.
