Research map / 3D vision / Depth and stereo
Elite360D: Towards Efficient 360 Depth Estimation via Semantic- and Distance-Aware Bi-Projection Fusion
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 Elite360D, a compact bi-projection fusion module (B2F) that fuses ERP features with an icosahedral point set (ICOSAP) via semantic- and distance-aware attention and gated fusion for efficient 360 depth estimation.
Problem
Monocular 360 depth estimation suffers from limited local receptive fields of ERP features and the high cost of existing cross-projection (bi-projection) fusion methods.
Approach
Replace expensive bi-projection fusion with a compact B2F module that attends from ERP pixels to an icosahedral point set (ICOSAP) using both semantic distance (query-key) and spatial distance (coordinate) embeddings, then adaptively fuses local and global features with a gated block.
Key results
- On Matterport3D with ResNet-34, Elite360D reaches Abs Rel 0.1115 and RMSE 0.4875, outperforming UniFuse by 2.53% in Abs Rel with only about 25.5M params.
- On Structured3D with ResNet-34, Elite360D improves Sq Rel by 4.48% and delta1 by 1.99% over UniFuse.
- Adds only ~1M parameters over the ERP baseline (C=64) while improving Abs Rel by over 10% on Matterport3D with ResNet-18 and ResNet-34.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ImageNet pre-training to ERP encoder (ablation) | Elite360D · M3D | Abs Rel: 0.1596 -> 0.1115; 0.1211 -> 0.1048; 0.1515 -> 0.1423 | Helped |
| Replace B2F with SFA plus addition (ablation) | Elite360D · M3D | Abs Rel: 0.1115 -> 0.1276 | Hurt |
| Replace B2F with SFA plus concatenation (ablation) | Elite360D · M3D | Abs Rel: 0.1115 -> 0.1191 | Hurt |
| Remove distance-aware attention from B2F (ablation) | Elite360D · M3D | Abs Rel: 0.1115 -> 0.1204 | Hurt |
| Remove semantic-aware attention from B2F (ablation) | Elite360D · M3D | Abs Rel: 0.1115 -> 0.1184 | Hurt |
| Replace gated fusion with addition (ablation) | Elite360D · M3D | Abs Rel: 0.1115 -> 0.1685 | Hurt |
| Replace gated fusion with average fusion (ablation) | Elite360D · M3D | Abs Rel: 0.1115 -> 0.1198 | Hurt |
| Replace gated fusion with concatenation (ablation) | Elite360D · M3D | Abs Rel: 0.1115 -> 0.1145 | Hurt |
| Replace gated fusion with adaptive fusion (ablation) | Elite360D · M3D | Abs Rel: 0.1115 -> 0.1244 | Hurt |
| Replace ICOSAP with cubemap projection (ablation) | Elite360D · M3D | Abs Rel: 0.1272 -> 0.1369 | Hurt |
| Replace ICOSAP with tangent projection (N=18) (ablation) | Elite360D · M3D | Abs Rel: 0.1272 -> 0.1328 | Hurt |
| Vary ICOSAP point feature number to N=20 (ablation) | Elite360D · M3D | Abs Rel: 0.1115 -> 0.1157 | Hurt |
Surprising findings
- An ERP-only baseline with a pre-trained EfficientNet-B5 backbone outperforms existing specifically designed 360 depth methods, according to the authors.
- Performance on Stanford2D3D lags behind some baselines, which the authors attribute to limited data for the transformer-based point encoder.
Limitations
- Does not consistently beat prior bi-projection methods on Stanford2D3D.
- Results depend on backbone choice and ICOSAP point number N, with N=80 chosen empirically.
- Re-trained baselines under unified settings, so comparisons rely on the authors' re-implementation.
Experimental setup
Datasets: Matterport3D Stanford2D3D Structured3D
Benchmarks: Matterport3D Stanford2D3D Structured3D
Metrics: Abs Rel Sq Rel RMSE delta1 delta2 delta3 #Params #FLOPs
Baselines: EGFormer PanoFormer BiFuse BiFuse++ UniFuse OmniFusion HRDFuse
Builds on: UniFuse BiFuse HRDFuse Point Transformer
Closest papers by meaning
- Omnidirectional Image Super-resolution via Bi-projection Fusion (AAAI 2024 · similarity 0.85)
- RPG360: Robust 360 Depth Estimation with Perspective Foundation Models and Graph Optimization (NeurIPS 2025 · similarity 0.85)
- Depth Anywhere: Enhancing 360 Monocular Depth Estimation via Perspective Distillation and Unlabeled Data Augmentation (NeurIPS 2024 · similarity 0.84)
- Spherical Pseudo-Cylindrical Representation for Omnidirectional Image Super-resolution (AAAI 2024 · similarity 0.84)
- HELVIPAD: A Real-World Dataset for Omnidirectional Stereo Depth Estimation (CVPR 2025 · similarity 0.84)
- Multi-view Reconstruction via SfM-guided Monocular Depth Estimation (CVPR 2025 · similarity 0.83)
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.