Research map / 3D vision / Gaussian splatting
EGSRAL:An Enhanced 3D Gaussian Splatting Based Renderer with Automated Labeling for Large-Scale Driving Scene
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 EGSRAL, an enhanced Deformable 3D Gaussian Splatting renderer for large-scale driving scenes that uses only images, plus an MLP adaptor that maps dataset annotations to novel views for automatic 2D/3D labeling.
Problem
Novel view synthesis of large-scale driving scenes with 3D GS typically depends on extra data (depth, boxes, object trajectories) and produces unlabeled synthetic images, limiting use in downstream perception training.
Approach
Extends Deformable 3D GS with a per-Gaussian state attribute, a deformation enhancement module and a neural opacity module, adds a spatial group-based training/rendering strategy for large scenes, and trains an MLP adaptor (three pose constraints) to map original-world camera poses to SfM-world poses so that nuScenes 3D boxes can be transformed to novel views as labels.
Key results
- On nuScenes (DrivingGaussian scenes), EGSRAL reaches PSNR 29.04, SSIM 0.883, LPIPS 0.162 versus DrivingGaussian's 28.74 / 0.865 / 0.237, without extra annotations.
- On KITTI City, EGSRAL achieves PSNR 23.60 and SSIM 0.787 vs. Deformable 3D GS at 21.33 / 0.713; the ablation shows the grouping strategy adds +0.76 PSNR.
- Adding 3x adaptor-generated novel-view images raises Co-DETR sample-set mAP from 23.7 to 26.8 and MonoLSS AP3D (sample) from 17.21 to 20.31.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EGSRAL instead of READ | Novel view synthesis · KITTI City | PSNR: 23.48 -> 23.60 | Helped |
| Use EGSRAL instead of READ | Novel view synthesis · KITTI City | LPIPS: 0.158 -> 0.192 | Hurt |
| Use EGSRAL instead of READ | Novel view synthesis · KITTI City | SSIM: 0.787 -> 0.787 | No clear effect |
| Use EGSRAL instead of DrivingGaussian | Novel view synthesis · nuScenes-D | PSNR: 28.74 -> 29.04 | Helped |
| Use EGSRAL instead of DrivingGaussian | Novel view synthesis · nuScenes-D | LPIPS: 0.237 -> 0.162 | Helped |
| Use EGSRAL instead of Deformable 3D GS | Novel view synthesis · nuScenes-S | PSNR: 33.43 -> 34.43 | Helped |
| Add DEM (ablation) | Deformable 3D GS · KITTI City | PSNR: 21.33 -> 22.37 | Helped |
| Add OEM (ablation) | Deformable 3D GS · KITTI City | PSNR: 21.33 -> 22.32 | Helped |
| Add GPS to EGSRAL with DEM and OEM (ablation) | EGSRAL · KITTI City | PSNR: 22.84 -> 23.60 | Helped |
| Increase GPS group count from 4 to 8 (ablation) | EGSRAL GPS · KITTI City | PSNR: 22.64 -> 23.60 | Helped |
| Increase GPS group count from 8 to 10 (ablation) | EGSRAL GPS · KITTI City | PSNR: 23.60 -> 23.38 | Hurt |
| Increase synthesized dataset amount from 1× to 3× | Co-DETR · nuScenes, sample | mAP: 23.7 -> 26.8 | Helped |
Surprising findings
- The adaptor-based annotation transfer yields much better box accuracy than the Umeyama matrix alignment baseline (AP 72.53% vs 21.71%).
Limitations
- Reconstruction quality depends on the accuracy of SfM point clouds used for initialization.
- Weaker high-frequency ground texture reconstruction yields higher LPIPS on KITTI than READ.
- Adaptor training requires camera pose pairs from the dataset and was trained per scene (about 2 hours on AMD MI100).
Experimental setup
Datasets: KITTI nuScenes
Benchmarks: KITTI City nuScenes (DrivingGaussian scenes) nuScenes-S (S-NeRF scenes) nuScenes 2D/3D detection augmentation
Metrics: PSNR SSIM LPIPS mAP AP50 AP75 AP3D AP_BEV AP average distance
Baselines: NPBG ADOP Deepview MPI READ 3D GS Mip-splatting Deformable 3D GS Mip-NeRF Mip-NeRF360 Urban-NeRF S-NeRF
Builds on: Deformable 3D GS 3D Gaussian Splatting READ DrivingGaussian
Closest papers by meaning
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.90)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.90)
- GGS: Generalizable Gaussian Splatting for Lane Switching in Autonomous Driving (AAAI 2025 · similarity 0.90)
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.90)
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.89)
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.