OPV2V: An Open Benchmark Dataset and Fusion Pipeline for Perception with Vehicle-to-Vehicle Communication
DOI: 10.1109/icra46639.2022.9812038
archive: archived pipeline: cataloged verified
Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)
Summary
This paper introduces OPV2V, the first large-scale open benchmark dataset and fusion pipeline for cooperative perception using Vehicle-to-Vehicle (V2V) communication. The research addresses the critical challenge of occlusion in autonomous driving, where single-vehicle sensors fail to detect objects hidden by other vehicles or infrastructure. While V2V communication offers a solution by sharing sensing data among Connected Automated Vehicles (CAVs), the field has been hindered by the lack of a standardized, large-scale dataset for benchmarking algorithms. Existing datasets are either too small, lack ground truth, or are not publicly available, preventing rigorous evaluation of cooperative perception technologies. To address this gap, the authors constructed the OPV2V dataset using the CARLA simulator and the OpenCDA co-simulation framework. The dataset comprises 73 diverse driving scenes across eight CARLA towns and a digital twin of Culver City, Los Angeles, totaling 11,464 frames and 232,913 annotated 3D vehicle bounding boxes. Each CAV is equipped with four cameras, a 64-channel LiDAR, and GPS/IMU sensors. The dataset covers various road types, including intersections and ramps, with varying traffic densities and CAV configurations (2 to 7 vehicles per frame). Additionally, the authors proposed an Attentive Intermediate Fusion pipeline, which aggregates intermediate features from neighboring CAVs using self-attention mechanisms to capture spatial interactions while managing bandwidth constraints through feature compression. The study evaluated 16 models, combining four state-of-the-art LiDAR detectors (SECOND, VoxelNet, PIXOR, PointPillar) with three fusion strategies: early fusion (raw data), late fusion (detection outputs), and intermediate fusion (proposed pipeline). Results demonstrated that all fusion methods significantly outperformed single-vehicle perception, achieving over 10% gains in Average Precision (AP) at IoU 0.7. Intermediate fusion generally yielded the best performance, surpassing early fusion due to its ability to effectively model correlations between vehicles. Notably, the proposed pipeline maintained high accuracy even with a 4096x compression rate, dropping performance by only ~3% compared to uncompressed intermediate fusion, thus offering a robust trade-off between bandwidth efficiency and detection accuracy. Performance on the realistic Culver City data was lower than on simulated CARLA towns, highlighting a domain gap that underscores the need for real-world data integration. The significance of this work lies in providing the community with a comprehensive, reproducible benchmark for V2V perception, facilitating the development of robust cooperative driving systems. The OPV2V dataset and the Attentive Intermediate Fusion pipeline enable researchers to evaluate and improve algorithms for handling occlusions and sparse sensor data. By demonstrating that intermediate fusion can achieve state-of-the-art performance with minimal bandwidth requirements, the paper establishes a viable pathway for deploying cooperative perception in real-world vehicular networks. The release of the dataset, code, and benchmark tools aims to accelerate progress in this emerging field.
Provenance
The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | OpenAlex-citations | — | — | 1 | 2026-06-20 |
| archive | success | semantic_scholar | — | — | 6 | 2026-06-26 |
| extract | success | cached | — | — | 2 | 2026-06-26 |
| clean | success | clean | — | — | 1 | 2026-06-20 |
| chunk | success | chunk | — | — | 1 | 2026-06-20 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-20 |
| promote | success | — | — | — | 1 | 2026-06-20 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 1 | 2026-06-26 |
| tag | success | vector_similarity | — | — | 6 | 2026-06-20 |
| verify | success | — | — | — | 1 | 2026-06-26 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.
Information type
What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).
- Methodological Resource: tool software, dataset resource