Research map / 3D vision / Depth and stereo
HELVIPAD: A Real-World Dataset for Omnidirectional Stereo Depth Estimation
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.
Introduces HELVIPAD, a real-world 360° top-bottom stereo dataset with LiDAR-derived depth labels and depth-completion augmentation, and adapts IGEV-Stereo with polar angle input and circular padding (360-IGEV-Stereo) for omnidirectional stereo.
Problem
Omnidirectional stereo depth estimation is underexplored due to a lack of real-world 360° stereo data with pixel-wise depth labels, and standard stereo models struggle with equirectangular distortions.
Approach
Collect a top-bottom two-360°-camera plus LiDAR dataset, map LiDAR to equirectangular images, densify labels via temporal aggregation and interpolation-based depth completion, and adapt IGEV-Stereo with a polar angle map and circular padding.
Key results
- 360-IGEV-Stereo achieves the best depth MAE of 1.720 m and RMSE of 4.297 m on the HELVIPAD test set, versus 1.860 m MAE for IGEV-Stereo.
- Circular padding reduces LRCE from 1.153 to 0.388 for 360-IGEV-Stereo (ablation).
- Depth completion raises labeled pixel ratio from 12% to 61% and improves 360SD-Net and PSMNet depth MARE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Adapt IGEV-Stereo with polar angle input and circular padding | IGEV-Stereo · HELVIPAD test set | Disparity MAE (°): 0.225 -> 0.188 | Helped |
| Add circular padding (ablation) | 360-IGEV-Stereo · HELVIPAD test set | LRCE: 1.153 -> 0.388 | Helped |
| Add circular padding (ablation) | 360-IGEV-Stereo · HELVIPAD test set | Depth MARE: 0.130 -> 0.130 | No clear effect |
| Add photometric data augmentation (ablation) | 360-IGEV-Stereo · HELVIPAD test set | Depth MARE: 0.135 -> 0.130 | Helped |
| Train using depth-completed augmented labels (ablation) | 360SD-Net · HELVIPAD test set | Depth MARE: 0.17 -> 0.15 | Helped |
| Train on Indoor+Outdoor data rather than only a single environment (ablation) | Stereo depth estimation models · HELVIPAD cross-scene evaluation | Depth MARE | Helped |
| Train on indoor data and evaluate on outdoor scenes, particularly at night (ablation) | Stereo depth estimation models · HELVIPAD cross-scene evaluation | Depth MARE | Hurt |
Surprising findings
- Recent conventional stereo model IGEV-Stereo outperforms the older omnidirectional 360SD-Net on depth MAE and RMSE.
- Depth completion augmentation gives smaller gains for advanced models like IGEV-Stereo and 360-IGEV-Stereo.
Limitations
- Depth completion is limited for fast-moving objects such as cars due to temporal aggregation.
- Indoor-trained models generalize poorly to outdoor night scenes.
- Dataset collected on a single university campus with one rig configuration.
Experimental setup
Datasets: HELVIPAD
Benchmarks: HELVIPAD test set
Metrics: MAE RMSE MARE LRCE
Baselines: PSMNet 360SD-Net IGEV-Stereo
Builds on: IGEV-Stereo PSMNet 360SD-Net
Closest papers by meaning
- Depth Anywhere: Enhancing 360 Monocular Depth Estimation via Perspective Distillation and Unlabeled Data Augmentation (NeurIPS 2024 · similarity 0.87)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.87)
- DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images (NeurIPS 2026 · similarity 0.86)
- Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data (CVPR 2024 · similarity 0.86)
- RPG360: Robust 360 Depth Estimation with Perspective Foundation Models and Graph Optimization (NeurIPS 2025 · similarity 0.86)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · 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.