Research map / Detection and segmentation / 3D detection and driving
Generalizable Multi-Camera 3D Object Detection from a Single Source via Fourier Cross-View Learning
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 Fourier Cross-View Learning (FCVL), combining Fourier Hierarchical Augmentation and a Fourier cross-view semantic consistency loss, to improve single-source generalization of multi-camera BEV 3D object detectors.
Problem
Multi-camera 3D object detectors degrade under distribution shifts (weather, lighting, sensor corruptions) when trained on only a single source domain.
Approach
Uses frequency-domain augmentation (Frequency Jitter at image level, Amplitude Transfer at feature level) to diversify the single source domain, and a triplet loss on phase components across adjacent camera views to learn domain-invariant features; theoretical support via augmentation graph theory.
Key results
- On nuScenes-C, FCVL raises BEVFormer's average OoD NDS from 0.3028 to 0.3567 across eight corruption domains.
- FCVL improves BEVDet's OoD average NDS from 0.2017 to 0.2677 and BEVDepth's from 0.2365 to 0.2870 (PD-BEV-based comparison also reported).
- On Argoverse 2 with Far3D, OoD average mAP rises from 0.0970 to 0.1346.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FCVL | BEVFormer · nuScenes-C OoD Avg. (8 domains) | NDS: 0.3028 -> 0.3567 | Helped |
| Add FCVL | BEVDepth · nuScenes-C OoD Avg. (8 domains) | NDS: 0.2365 -> 0.2870 | Helped |
| Add FCVL | BEVDet · nuScenes-C OoD Avg. (8 domains) | NDS: 0.2017 -> 0.2677 | Helped |
| Add FCVL | Far3D · Argoverse 2 OoD Avg. | mAP: 0.0970 -> 0.1346 | Helped |
| Use amplitude jitter only (ablation) | BEVDet · nuScenes-C OoD Avg. | NDS: 0.2365 -> 0.2735 | Helped |
| Use phase jitter only (ablation) | BEVDet · nuScenes-C OoD Avg. | NDS: 0.2365 -> 0.2690 | Helped |
| Jitter both amplitude and phase (ablation) | BEVDet · nuScenes-C OoD Avg. | NDS: 0.2365 -> 0.2767 | Helped |
| Use Amplitude Transfer only (ablation) | BEVDet · nuScenes-C OoD Avg. | NDS: 0.2365 -> 0.2570 | Helped |
| Combine Frequency Jitter and Amplitude Transfer (ablation) | BEVDet · nuScenes-C OoD Avg. | NDS: 0.2365 -> 0.2843 | Helped |
| Add Fourier Cross-View Semantic Consistency Loss to FHiAug (ablation) | BEVDet · nuScenes-C OoD Avg. | NDS: 0.2843 -> 0.2870 | Helped |
| Insert Amplitude Transfer at positions 0–3 rather than 0–5 (ablation) | BEVFormer · nuScenes-C OoD Avg. | NDS: 0.3267 -> 0.3294 | Helped |
| Increase the cross-view loss weight from γ = 0.5 to γ = 1.0 (ablation) | BEVDet · nuScenes-C OoD Avg. | NDS: 0.2024 -> 0.2296 | Helped |
Surprising findings
- Common 2D augmentations such as geometric transforms and style transfer are argued to be ill-suited to BEV detection, and the paper reports that a cross-perturbation baseline (CPerb) gives little or no gain on BEVDepth and BEVFormer.
Limitations
- Several hyperparameters (eta, lambda, p_i, p_f, gamma) must be tuned.
- Snow performance remains much lower than other corruption types despite gains.
- Evaluated only on synthetic corruptions and a single daylight-to-night real-world case.
Experimental setup
Datasets: nuScenes nuScenes-C Argoverse 2
Benchmarks: nuScenes nuScenes-C Argoverse 2
Metrics: NDS mAP
Baselines: CPerb DSU DAC-SC FACT PD-BEV CPerb on BEVDepth
Builds on: BEVDepth BEVFormer BEVDet FACT CPerb DSU
Closest papers by meaning
- Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection (NeurIPS 2024 · similarity 0.88)
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.88)
- Geometry-Guided Domain Generalization for Monocular 3D Object Detection (AAAI 2024 · similarity 0.87)
- Generalizable Fourier Augmentation for Unsupervised Video Object Segmentation (AAAI 2024 · similarity 0.86)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.86)
- CorrBEV: Multi-View 3D Object Detection by Correlation Learning with Multi-modal Prototypes (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.