Research map / 3D vision / Gaussian splatting

FlexDrive: Toward Trajectory Flexibility in Driving Scene Gaussian Splatting Reconstruction and Rendering

CVPR 2025 · 3D vision › Gaussian splatting · approach: Compact and efficient Gaussian splatting · 1 citations (OpenAlex)

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 FlexDrive, which uses Inverse View Warping with a Depth Bootstrap strategy to supervise out-of-path viewpoints in 3DGS driving-scene reconstruction, and introduces a CARLA-based out-of-path benchmark.

Problem

Driving scene Gaussian splatting reconstructions render well along pre-recorded vehicle paths but degrade for out-of-path viewpoints because no ground-truth images exist for those views.

Approach

Inverse View Warping renders the in-path image at a sampled out-of-path viewpoint using occlusion-aware rasterization and then rearranges pixels back to the in-path layout, so the real in-path image serves as supervision; dense depth for lifting pixels comes from a Depth Bootstrap that rectifies rendered depth with LiDAR via a linear fit.

Key results

• On the CARLA-based out-of-path benchmark, FlexDrive reaches 26.23 PSNR and 0.372 LPIPS versus 24.84 PSNR for LidaRF and 21.65 for PVG.
• On Waymo in-path rendering, FlexDrive gets 31.25 PSNR on static and 30.76 on dynamic scenes, comparable to StreetGaussian (31.35 / 30.73).
• Ablation on LidaRF-GS baseline: adding DB and IVW raises CARLA out-of-path PSNR from 24.79 to 26.23.

Measured effects reported in the paper

Surprising findings

• Without occlusion-aware rasterization (beta=0), single-scene PSNR collapses to 20.12 versus 32.23 at beta=0.95, showing occlusion handling is critical.

Limitations

• Out-of-path evaluation on Waymo relies on FID, which the authors note is an unreliable quality indicator.
• Out-of-path benchmark is built on the CARLA simulator rather than real driving data.
• Method is restricted to views near the recorded trajectory; completely free camera movement is left to future work.

Experimental setup

Datasets: Waymo Open Dataset CARLA simulator data

Benchmarks: Waymo Open Dataset CARLA-based out-of-path benchmark

Metrics: PSNR SSIM LPIPS FID

Baselines: 3D GS PVG EmerNeRF LidaRF NeuraD StreetGaussian LidaRF-GS

Builds on: 3D Gaussian Splatting LidaRF PVG StreetGaussian

Closest papers by meaning

• DrivingForward: Feed-forward 3D Gaussian Splatting for Driving Scene Reconstruction from Flexible Surround-view Input (AAAI 2025 · similarity 0.89)
• SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving (CVPR 2025 · similarity 0.89)
• DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes (CVPR 2024 · similarity 0.88)
• EGSRAL:An Enhanced 3D Gaussian Splatting Based Renderer with Automated Labeling for Large-Scale Driving Scene (AAAI 2025 · similarity 0.88)
• Swift4D: Adaptive divide-and-conquer Gaussian Splatting for compact and efficient reconstruction of dynamic scene (ICLR 2025 · similarity 0.87)
• Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.87)