Research map / Vision-language and multimodal / CLIP and zero-shot
ROADWork: A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones
Anurag Ghosh, Shen Zheng, Robert Tamburo, Khiem Vuong, Juan Alvarez-Padilla, Hailiang Zhu, Michael Cardei, Nicholas Dunn, Christoph Mertz, Srinivasa G. Narasimhan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune Mask DINO on ROADWork instead of open-vocabulary pretrained | Mask DINO · ROADWork work zone object segmentation | AP: +32.2 AP | Improved |
| add SAM2 video label propagation (ablation) | Mask DINO · ROADWork work zone object segmentation | AP: 36.4 -> 39.0 | Improved |
| use manual masks instead of SAM pseudo masks (ablation) | Mask R-CNN · ROADWork pseudo-mask supervision | AP: +5.2 AP | Improved |
| add 5 or 10 extra points to box prompts for SAM (ablation) | SAM pseudo masks · ROADWork pseudo-mask supervision | AP: 24.7 -> 25.8 | Improved |
| use SAM2 instead of SAM for pseudo masks (ablation) | pseudo-mask supervision · ROADWork pseudo-mask supervision | AP: 24.7 -> 26.4 | No clear effect |
| add copy-paste augmentation (ablation) | Mask R-CNN · ROADWork fine-grained sign segmentation | AP: 17.6 -> 25.0 | Improved |
| add video label propagation (ablation) | Mask DINO · ROADWork fine-grained sign segmentation | AP: 30.0 -> 32.8 | Improved |
| add crop-rescaling with detector before Glass (ablation) | Glass · ROADWork sign text reading (all signs) | 1-NED: 65.4 -> 79.6 | Improved |
| increase crop scale from 1x to 3x (ablation) | Glass · ROADWork sign text reading (all signs) | Word Acc.: 56.3 -> 59.6 | Improved |
| add crop-rescaling for small signs (ablation) | Glass · ROADWork sign text reading (small signs) | Word Acc.: 13.2 -> 32.3 | Improved |
| fine-tune LLaVA-1.5-7B on ROADWork | LLaVA-1.5-7B · ROADWork scene description | SPICE: 9.9 -> 42.7 | Improved |
| add predicted object context to prompt (ablation) | LLaVA-1.5-7B · ROADWork scene description | SPICE: 42.7 -> 46.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- OMG-LLaVA — OMG-LLaVA: Bridging Image-level, Object-level, Pixel-level Reasoning and Understanding (NeurIPS 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- RoadBench: Benchmarking MLLMs on Fine-Grained Spatial Understanding and Reasoning under Urban Road Scenarios (ECCV 2026 · similarity 0.85)
- RoadSceneBench: A Lightweight Benchmark for Mid-Level Road Scene Understanding (CVPR 2026 · similarity 0.84)
- UrbanIng-V2X: A Large-Scale Multi-Vehicle, Multi-Infrastructure Dataset Across Multiple Intersections for Cooperative Perception (NeurIPS 2025 · similarity 0.84)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.84)
- Rethinking Lanes and Points in Complex Scenarios for Monocular 3D Lane Detection (CVPR 2025 · similarity 0.82)
- SpaceVista: All-Scale Visual Spatial Reasoning from mm to km (ICML 2026 · similarity 0.82)
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.
