Research map / 3D vision / Gaussian splatting
Toward Real-world BEV Perception: Depth Uncertainty Estimation via Gaussian Splatting
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 GaussianLSS, a 2D unprojection BEV perception method that models per-pixel depth uncertainty as the variance of the depth distribution and splats 3D Gaussians into BEV features, achieving state-of-the-art among unprojection methods on nuScenes.
Problem
BEV perception for autonomous driving: projection-based methods are accurate but computationally expensive, while LSS-style unprojection methods are fast but sensitive to depth errors and lack explicit uncertainty modeling.
Approach
Treat the LSS per-pixel depth distribution as a Gaussian: compute its mean and variance, convert it to a 3D Gaussian in the ego frame, and rasterize these Gaussians (with learned opacity) onto multi-scale BEV planes via Gaussian Splatting, so object extent is captured implicitly by the depth spread.
Key results
- GaussianLSS reaches 38.0 vehicle IoU at 224x480 without visibility filtering, outperforming unprojection baselines such as PointBEV-free methods FIERY (35.8) and BAEFormer (36.0).
- Runs at 67.9 FPS with 0.36 GiB memory, about 2x faster than PointBEV (32.0 FPS) and roughly 0.3x its memory (1.26 GiB), with 0.7 IoU lower than PointBEV (38.7).
- Achieves 46.1 vehicle IoU with visibility filtering at 448x800, versus PointBEV at 47.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GaussianLSS instead of PointBeV | nuScenes vehicle segmentation, no visibility filtering, 224 × 480 · nuScenes vehicle segmentation, no visibility filtering, 224 × 480 | IoU: 38.7 -> 38.0 | Hurt |
| use GaussianLSS instead of PointBeV | nuScenes vehicle segmentation, no visibility filtering, 448 × 800 · nuScenes vehicle segmentation, no visibility filtering, 448 × 800 | IoU: 42.1 -> 40.6 | Hurt |
| use GaussianLSS instead of PointBeV | nuScenes vehicle segmentation, visibility filtering, 224 × 480 · nuScenes vehicle segmentation, visibility filtering, 224 × 480 | IoU: 44.0 -> 42.2 | Hurt |
| use GaussianLSS instead of PointBeV | nuScenes vehicle segmentation, visibility filtering, 448 × 800 · nuScenes vehicle segmentation, visibility filtering, 448 × 800 | IoU: 47.6 -> 46.1 | Hurt |
| use GaussianLSS instead of PointBeV | nuScenes pedestrian segmentation · nuScenes pedestrian segmentation, visibility filtering, 224 × 480 | IoU: 18.5 -> 17.4 | Hurt |
| use GaussianLSS instead of PointBeV | nuScenes vehicle segmentation · nuScenes vehicle segmentation, 224 × 480 | FPS: 32.0 -> 67.9 | Helped |
| use GaussianLSS instead of PointBeV | nuScenes vehicle segmentation · nuScenes vehicle segmentation, 224 × 480 | Mem GiB: 1.26 -> 0.36 | Helped |
| replace depth uncertainty modeling with directly predicting spatial extent (ablation) | GaussianLSS · nuScenes vehicle segmentation | IoU: 1.0% lower performance | Hurt |
| increase the error tolerance coefficient k beyond 1.25 (ablation) | GaussianLSS · nuScenes vehicle segmentation | IoU | Hurt |
| set the error tolerance coefficient k between 0.5 and 1.25 (ablation) | GaussianLSS · nuScenes vehicle segmentation | IoU | No clear effect |
Surprising findings
- Directly predicting a fixed spatial extent instead of modeling uncertainty lowers performance by 1.0% IoU.
- Roughly 80% of Gaussians end up with opacity below 0.01 after training, so most features are filtered out.
Limitations
- Performance on fine object shape is poorer than projection-based methods.
- Vehicle IoU remains below PointBEV at the same resolution (0.7% gap).
- Evaluated only in single-frame setting; temporal modeling left for future work.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes
Metrics: IoU FPS memory (GiB)
Baselines: BEVFormer SimpleBEV PointBEV FIERY CVT LaRa BAEFormer LSS ST-P3
Builds on: Lift-Splat-Shoot Gaussian Splatting PointBEV Focal loss
Closest papers by meaning
- A Black-Box Evaluation Framework for Semantic Robustness in Bird’s Eye View Detection (AAAI 2025 · similarity 0.88)
- Predictive Uncertainty Quantification for Bird's Eye View Segmentation: A Benchmark and Novel Loss Function (ICLR 2025 · similarity 0.88)
- M-BEV: Masked BEV Perception for Robust Autonomous Driving (AAAI 2024 · similarity 0.87)
- DiffBEV: Conditional Diffusion Model for Bird’s Eye View Perception (AAAI 2024 · similarity 0.87)
- Improving Bird's Eye View Semantic Segmentation by Task Decomposition (CVPR 2024 · similarity 0.87)
- Instance-Aware Multi-Camera 3D Object Detection with Structural Priors Mining and Self-Boosting Learning (AAAI 2024 · similarity 0.87)
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.