Research map / 3D vision / 3D generation
RoScenes: A Large-scale Multi-view 3D Dataset for Roadside Perception
Xiaosu Zhu, Hualian Sheng, Sijia Cai, Bing Deng, Shaopeng Yang, Qiao Liang, Ken Chen, Lianli Gao, Jingkuan Song, Jieping Ye
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 |
|---|---|---|---|
| use feature-guided position embedding (RoBEV) instead of static position embedding (PETRv2) | RoBEV · RoScenes val avg | NDS: 0.627 -> 0.713 | Improved |
| use learnable shared position encoding instead of sinusoidal (ablation) | RoBEV · RoScenes val avg | NDS: 0.608 -> 0.713 | Improved |
| share position encoding network between queries and points instead of separate networks (ablation) | RoBEV · RoScenes val avg | NDS: 0.685 -> 0.713 | Improved |
| use implicit BEV methods instead of explicit BEV methods | BEV detectors · RoScenes val | NDS: 19.7% higher on average | Improved |
| use explicit BEV methods (training time and memory) | BEV detectors · RoScenes | training time and inference memory: 1.75x training time, 1.50x inference memory | Hurt |
| use nuScenes-pretrained VoVNet instead of ImageNet-pretrained (ablation) | RoBEV · RoScenes val | NDS: 0.6% drop for ImageNet pretraining | Improved |
| scale backbone to Swin-B, ViT-L or InternImage-XL (ablation) | RoBEV · RoScenes val avg | NDS: 0.713 -> 0.719 | No clear effect |
| train on all scenes instead of a single scene and test on all scenes (ablation) | RoBEV · RoScenes val easy | NDS: 0.396 -> 0.753 | Improved |
| apply image registration to perturbed images (ablation) | RoBEV · pRoScenes easy | NDS: 0.441 -> 0.745 | Improved |
| use RoBEV on single-layout nuScenes | RoBEV · nuScenes | NDS: 0.503 -> 0.505 | No clear effect |
| increase detection queries from 450 to 900 (ablation) | PETRv2 and RoBEV · RoScenes val | NDS: 7.7% average improvement | Improved |
| evaluate trained BEV methods on unseen scenes | all BEV methods · RoScenes zero-shot test | NDS: 0.713 -> 0.308 (RoBEV) | Hurt |
Closest papers by meaning
- BEVSpread: Spread Voxel Pooling for Bird's-Eye-View Representation in Vision-based Roadside 3D Object Detection (CVPR 2024 · similarity 0.88)
- MagicDrive: Street View Generation with Diverse 3D Geometry Control (ICLR 2024 · similarity 0.87)
- SOGDet: Semantic-Occupancy Guided Multi-View 3D Object Detection (AAAI 2024 · similarity 0.87)
- RCooper: A Real-world Large-scale Dataset for Roadside Cooperative Perception (CVPR 2024 · similarity 0.86)
- AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios (NeurIPS 2025 · similarity 0.86)
- UrbanIng-V2X: A Large-Scale Multi-Vehicle, Multi-Infrastructure Dataset Across Multiple Intersections for Cooperative Perception (NeurIPS 2025 · similarity 0.86)
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.
