Research map / 3D vision / 3D generation
HoloVIC: Large-scale Dataset and Benchmark for Multi-Sensor Holographic Intersection and Vehicle-Infrastructure Cooperative
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 HoloVIC, a 100k-frame real-world multi-sensor holographic intersection and vehicle-infrastructure cooperation dataset with 3D/2D annotations, global IDs, and benchmarks for detection, tracking, and VIC perception.
Problem
Single-view roadside perception suffers from occlusion and blind spots, and existing real-world VIC datasets use a single viewpoint with few sensors and limited annotated frames.
Approach
Builds holographic intersections with 6-18 overlapping roadside sensors (cameras, fisheyes, lidars) in four layouts plus synchronized vehicle-side data, annotates 3D boxes with track and global IDs linking devices across vehicle and roadside, and defines tasks and baselines on top.
Key results
- Lidar 3D detection with Transfusion-L reaches 45.51 3D AP on pedestrians at IoU 0.5 and PVRCNN 82.16 BEV AP on vehicles, while Mono3D PGD reaches 52.54 3D AP on vehicles.
- Multi-target tracking with FairMOT achieves 68.17 IDF1 and 63.35 MOTA on 2D MOT validation.
- VIC cooperation with roadside 4C+2L improves vehicle-side BEV AP beyond 50m from 46.95 to 92.31 range-level, and gives 8.9% IDF1 and 14.55% MOTA gains for the vehicle side.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add roadside perception to vehicle-only perception | Vehicle-side perception · VIC-1 validation, 50–70m | bev AP: 46.95 -> 59.82 | Helped |
| increase roadside sensors from 1C+1L to 4C+2L (ablation) | VIC perception · VIC-1 validation, 50–70m | bev AP: 59.82 -> 85.75 | Helped |
| add roadside perception to vehicle-only perception | Vehicle-side tracking · VIC-1 validation | IDF1: 86.57 -> 90.02 | Helped |
| increase roadside sensors from 1C+1L to 4C+2L (ablation) | VIC perception · VIC-1 validation | IDF1: 90.02 -> 95.47 | Helped |
| increase the number of sensors from 1C to 4C+2L (ablation) | Multi-sensor tracking · Int-1 validation | IDF1: 48.52 -> 77.81 | Helped |
| use 2C instead of 1C+1L (ablation) | Multi-sensor tracking · Int-1 validation | IDF1: 53.89 -> 63.17 | Helped |
| use Lidar 3D detection rather than Mono3D detection | Single-sensor perception · HoloVIC validation, cyclist and pedestrian detection | 3d AP | Helped |
| use image-based perception rather than Lidar 3D perception | Single-sensor perception · HoloVIC validation | AOS | Helped |
Surprising findings
- Adding more roadside cameras improves multi-sensor tracking monotonically, with 4C+2L far outperforming 2C+2L in MOTA and IDF1.
- Monocular 3D image-based detection has higher orientation AOS than Lidar 3D detection despite lower AP.
Limitations
- Benchmarks are reported on validation sets only; the test set has no labels and requires online submission.
- Only five intersections are used, and the VIC data covers one vehicle trajectory per collection.
Experimental setup
Datasets: HoloVIC
Benchmarks: HoloVIC
Metrics: mAP mAOS MOTA IDF1 MT ML
Baselines: FCOS3D PGD SECOND Pointpillars PVRCNN Transfusion-L DeepSort Tracktor FairMOT AB3DMOT
Builds on: DAIR-V2X AB3DMOT FCOS3D
Closest papers by meaning
- UrbanIng-V2X: A Large-Scale Multi-Vehicle, Multi-Infrastructure Dataset Across Multiple Intersections for Cooperative Perception (NeurIPS 2025 · similarity 0.90)
- TUMTraf V2X Cooperative Perception Dataset (CVPR 2024 · similarity 0.86)
- V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative Perception (NeurIPS 2025 · similarity 0.86)
- RCooper: A Real-world Large-scale Dataset for Roadside Cooperative Perception (CVPR 2024 · similarity 0.86)
- BEVSync: Asynchronous Data Alignment for Camera-based Vehicle-Infrastructure Cooperative Perception Under Uncertain Delays (AAAI 2025 · similarity 0.85)
- AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios (NeurIPS 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.