Research map / 3D vision / Depth and stereo
BevSplat: Resolving Height Ambiguity via Feature-Based Gaussian Primitives for Weakly-Supervised Cross-View Localization
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 BevSplat, which lifts each ground-image pixel to feature-bearing 3D Gaussians and renders them into a BEV feature map for weakly supervised cross-view pose estimation on KITTI and VIGOR.
Problem
Weakly supervised cross-view localization with noisy ground-truth poses, where height ambiguity in lifting ground images to BEV makes IPM distort objects and cross-view transformers are costly and hard to train with weak labels.
Approach
Each ground-image pixel is represented as N_p=3 feature-bearing 3D Gaussians (depth-initialized, with learned offsets, scales, rotations, opacities) that are alpha-blended into a BEV feature and confidence map, replacing IPM or cross-view transformers.
Key results
- On KITTI Same-Area with weak supervision (lambda1=1), BevSplat reaches 2.87 m mean localization error vs 6.81 m for G2SWeakly (lambda1=1, same setting listed).
- On VIGOR Cross-Area aligned-orientation with lambda1=1, BevSplat achieves 2.84 m mean error vs 4.70 m for G2SWeakly.
- On KITTI Cross-Area with lambda1=1, BevSplat achieves 6.20 m mean and 2.51 m median error, outperforming listed supervised methods in mean and median error.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace IPM-based BEV synthesis with feature-based Gaussian splatting | BevSplat on KITTI · KITTI Cross-Area, λ₁=1 | localization mean error (m): 12.15 -> 6.20 | Helped |
| replace G2SWeakly with BevSplat | VIGOR · VIGOR Cross-Area Aligned-orientation, λ₁=1 | mean error (m): 4.70 -> 2.84 | Helped |
| add optimizable Gaussian parameters (ablation) | BevSplat · KITTI Same-Area, λ₁=0 | mean error (m): 7.42 -> 5.82 | Helped |
| add optimizable Gaussian parameters (ablation) | BevSplat · KITTI Cross-Area, λ₁=1 | mean error (m): 7.86 -> 6.20 | Helped |
| use DPT fine-tuning instead of direct end-to-end training (ablation) | DINOv2(FiT) backbone on KITTI · KITTI Same-Area, λ₁=0 | mean error (m): 17.74 -> 5.82 | Helped |
| use DPT fine-tuning instead of LoRA (ablation) | DINOv2(FiT) backbone on KITTI · KITTI Same-Area, λ₁=0 | mean error (m): 16.29 -> 5.82 | Helped |
| replace DINOv2(FiT) with VGG (ablation) | BevSplat on KITTI · KITTI Cross-Area, λ₁=1 | mean error (m): 6.20 -> 8.02 | Hurt |
| replace DINOv2(FiT) with DINOv1 (ablation) | BevSplat on KITTI · KITTI Cross-Area, λ₁=1 | mean error (m): 6.20 -> 7.12 | Hurt |
| replace DINOv2(FiT) with original DINOv2 (ablation) | BevSplat on KITTI · KITTI Cross-Area, λ₁=1 | mean error (m): 6.20 -> 7.01 | Hurt |
| replace DepthAnythingV2 with DepthAnythingV1 (ablation) | BevSplat on KITTI · KITTI Same-Area, λ₁=1 | mean error (m): 2.87 -> 2.97 | Hurt |
| use 1 rather than 3 Gaussian primitives per pixel (ablation) | BevSplat on VIGOR · VIGOR Same-Area, λ₁=0 | mean error (m): 2.96 -> 3.05 | Hurt |
| use 4 rather than 3 Gaussian primitives per pixel (ablation) | BevSplat on VIGOR · VIGOR Same-Area, λ₁=0 | mean error (m): 2.96 -> 3.03 | Hurt |
Surprising findings
- A fully supervised variant of BevSplat performed worse than its weakly supervised (lambda1=1) version in the cross-area KITTI setting, which the authors attribute to overfitting to training-domain biases.
- Lift-Splat-Shoot and OrienterNet style depth-based re-sampling performed worse than IPM in several KITTI settings despite using full vertical information.
Limitations
- Inference speed is limited by 3DGS rendering overhead; BEV rendering takes 14 ms vs 4 ms for IPM.
- Performance degrades on long straight road segments with few distinctive visual features, causing ambiguous matches.
- Code and data are promised for future release only; no statistical significance or error bars are reported.
Experimental setup
Datasets: KITTI VIGOR
Benchmarks: KITTI VIGOR
Metrics: mean localization error (m) median localization error (m) recall at 1m/3m lateral and longitudinal azimuth error
Baselines: G2SWeakly IPM Lift-Splat-Shoot OrienterNet Boosting VFA CCVPE HC-Net DenseFlow LoRA fine-tuning
Builds on: 3D Gaussian Splatting G2SWeakly DINOv2 DPT FiT
Closest papers by meaning
- FG^2: Fine-Grained Cross-View Localization by Fine-Grained Feature Matching (CVPR 2025 · similarity 0.88)
- View From Above: Orthogonal-View aware Cross-view Localization (CVPR 2024 · similarity 0.88)
- Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images (NeurIPS 2026 · similarity 0.87)
- Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery (NeurIPS 2026 · similarity 0.86)
- Semantic Ambiguity Modeling and Propagation for Fine-Grained Visual Cross View Geo-Localization (AAAI 2025 · similarity 0.85)
- From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · 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.