Research map / Image generation / Text-to-image diffusion
Controllable Road Marking Generation
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.
Introduces controllable road-marking generation: a text-conditioned latent rectified-flow DiT fills center-region lane/road dividers and pedestrian crossings from a drivable-area mask and outer-ring markings, with blur-trained targets and a training-free Structured Gaussian Render post-process.
Problem
Road markings (lane dividers, road dividers, crosswalks) are authored manually, limiting quantitative analysis and scenario testing; existing methods mostly extract markings from observations rather than generating missing layouts from geometry and text.
Approach
Combines a latent rectified-flow DiT with a topology-aware clDice-style loss, Gaussian-blurred training targets, and a deterministic SGR post-process that fits Bezier curves to divider skeletons and re-renders them as anisotropic super-Gaussian primitives to sharpen thin markings.
Key results
- On AV2 test, the full system reaches 80.8 Buffered F1 and 50.2 clDice vs 38.8 and 24.6 for adapted LSR-DM.
- On Waymo, the same pipeline reaches 88.0 Buffered F1 and 66.2 clDice.
- Crosswalk deletion edits succeed in 75.1% of eligible tiles, but addition succeeds only in 49.6%, and single-to-double yellow editing in 31.0%.
Surprising findings
- Stronger text guidance increases edit success but also increases false activation on non-target structures, so target success does not ensure preservation.
- Enabling SGR does not uniformly improve Buffered F1 in the ablation (AV2 without topology loss: SGR gives 81.8 vs 80.5 for no-SGR, but with topology loss full model gives 80.8 vs 78.2 without SGR, while Waymo no-topology no-SGR has the highest F1 at 88.6).
Limitations
- Prompts are template-derived from metadata and fine-grained divider editing remains unreliable.
- Single-reference metrics do not capture the multiple plausible layouts; traffic-rule compliance and downstream planning benefits are not validated.
- Only three marking channels and a narrow external baseline set (single adapted baseline).
Experimental setup
Datasets: Argoverse 2 Waymo Open Dataset nuScenes-SG
Benchmarks: Argoverse 2 test (4,597 tiles) Waymo test (1,000 tiles) nuScenes-SG (261 tiles)
Metrics: Buffered F1 clDice success false activation fa-MSE delta target change
Baselines: LSR-DM Ours w/ CNN deblur
Builds on: rectified flow latent diffusion LSR-DM clDice Bezier Splatting
Closest papers by meaning
- Lane2Seq: Towards Unified Lane Detection via Sequence Generation (CVPR 2024 · similarity 0.86)
- RoadSceneBench: A Lightweight Benchmark for Mid-Level Road Scene Understanding (CVPR 2026 · similarity 0.86)
- DriveCombo: Benchmarking Compositional Traffic Rule Reasoning in Autonomous Driving (CVPR 2026 · similarity 0.86)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.86)
- Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map (CVPR 2025 · similarity 0.85)
- T2SG: Traffic Topology Scene Graph for Topology Reasoning in Autonomous Driving (CVPR 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.