CoVeRaP: Cooperative Vehicular Radar Perception Benchmark
- The paper presents a benchmark for cooperative vehicular perception that leverages mmWave FMCW radars to improve multi-vehicle 3-D detection performance.
- It introduces a unified pipeline supporting both middle and late fusion with a PointNet-style encoder that integrates spatial, Doppler, and intensity cues.
- The framework demonstrates significant detection gains under strict IoU thresholds, validating the benefits of cooperative radar sharing over single-vehicle setups.
CoVeRaP denotes “Cooperative Vehicular Perception through mmWave FMCW Radars”, a radar-centric cooperative-perception dataset and baseline framework for multi-vehicle 3-D object detection. It releases a 21,116-frame benchmark with time-aligned radar, camera, and GPS streams from multiple vehicles, and pairs that benchmark with a unified cooperative-perception pipeline supporting both middle fusion and late fusion. The reported baseline uses a multi-branch PointNet-style encoder with self-attention to combine spatial, Doppler, and intensity cues, then decodes 3-D bounding boxes and per-point depth confidence. The central claim is that cooperative FMCW-radar sharing materially improves detection robustness relative to single-vehicle operation, especially at strict IoU thresholds (Song et al., 22 Aug 2025).
1. Scope and nomenclature
In recent arXiv usage, the acronym CoVeRaP is overloaded. One 2025 work uses it for “Cooperative Vehicular Perception through mmWave FMCW Radars” (Song et al., 22 Aug 2025), while another uses it for “Coordinated Robustness Evaluation Framework for Vision-LLMs” (Babu et al., 5 Jun 2025). In radar-perception contexts, CoVeRaP refers to the former: a cooperative vehicular perception system and dataset organized around mmWave FMCW radars, V2V radar sharing, and radar-only 3-D detection.
This distinction matters because the vehicular CoVeRaP paper is not an adversarial-robustness framework for VLMs, but a benchmark and modeling pipeline for cooperative perception. Its problem setting is multi-vehicle scene understanding under sparse and noisy radar sensing, rather than multimodal representation attacks.
2. Problem formulation and motivation
The motivating premise is that automotive FMCW radars remain reliable in rain, fog, and glare, but their point clouds are sparse, noisy, and limited in viewpoint. A single radar typically observes only a fragment of the environment. CoVeRaP therefore studies cooperative perception in which vehicles share radar data to combine complementary viewpoints such as rear and side views, produce denser and more complete point clouds, reduce blind spots, and improve 3D object detection accuracy (Song et al., 22 Aug 2025).
The work is explicitly positioned against a gap in the literature: the lack of a large-scale, reproducible, multi-vehicle FMCW radar dataset for benchmarking cooperative radar perception. This suggests that the contribution is simultaneously infrastructural and algorithmic. The dataset provides synchronized multi-vehicle sensing, while the baseline framework operationalizes two canonical fusion paradigms—feature-level and prediction-level fusion—under a common experimental setup.
A recurrent misconception is that radar robustness to adverse weather is sufficient for high-quality 3-D perception by itself. The CoVeRaP results instead frame robustness as conditional on overcoming radar sparsity and viewpoint incompleteness through cooperation. The reported gains are therefore not attributed to radar alone, but to cooperative radar sharing and fusion design.
3. Dataset design, instrumentation, and annotation
CoVeRaP is collected with two vehicles—an ego vehicle and an assistant vehicle—in a controlled environment (campus parking lot). Each vehicle carries the same three sensing components: a TI AWR1843 BOOST FMCW radar, a Google Pixel 6a RGB camera, and a Sparkfun NEO-M8P-2 GPS-RTK unit with centimeter-level accuracy and real-time correction (Song et al., 22 Aug 2025).
| Component | Configuration | Role |
|---|---|---|
| FMCW radar | TI AWR1843 BOOST, 77 GHz, 3 TX / 4 RX | Radar point-cloud sensing |
| RGB camera | Google Pixel 6a | Ground-truth annotation and temporal/spatial alignment |
| GPS-RTK | Sparkfun NEO-M8P-2 | Cross-vehicle spatial alignment |
The radar configuration is specified as 77 GHz center frequency, 30 MHz/μs chirp sweep rate, 256 samples per chirp, 60 frames captured, 100 ms duration per frame, 10,000 ksps ADC sampling rate, 15° azimuth resolution, and 60° elevation resolution. The dataset modalities comprise 3D radar point clouds with , velocity, range, bearing, intensity per point, synchronized RGB images, GPS coordinates, and 3D bounding box annotations.
The collection protocol comprises 100 scenarios × 600 frames each (60 seconds per scenario), from which 11 scenes, 21,116 frames are sampled for benchmarking. Synchronization is handled at two levels. Intra-vehicle alignment uses radar at 10 Hz, camera alignment by closest timestamp, and GPS at 1 Hz, interpolated to 1000 Hz. Inter-vehicle alignment is event-based (motion-onset) alignment, cross-checked by camera frames. The benchmark further provides four dataset variants: Rear-View Only, Side-View Only, Fused Ego-Rear, and Fused Ego-Side.
Point-cloud preparation is deliberately aggressive. The pipeline keeps the top 128 points by Doppler-FFT, filters multipath, and retains the top 20% by intensity for robustness and stability. For training, the baseline uses top 70 points by intensity retained for each radar. Ground-truth 3D boxes are generated by ground-truth interpolation and validation with images, and are described as interpolated by phase and verified visually via image/radar overlays.
4. Cooperative-perception framework and baseline architecture
The CoVeRaP framework supports two fusion strategies. In middle fusion, each vehicle first extracts features from its radar point cloud; the assistant vehicle’s features are then shifted into the ego’s coordinate frame using GPS offsets, and the fused feature set is used for joint detection. In late fusion, each vehicle independently predicts 3-D bounding boxes in its own frame; the assistant vehicle’s boxes are transformed into the ego frame, and a shallow neural net fuses the predicted boxes and confidence scores (Song et al., 22 Aug 2025).
The baseline detector is a PointNet-style multi-branch feature encoder. It has three parallel branches:
- Spatial position branch using MLPs to encode into 256-dim features.
- Dynamics branch encoding velocity, range, bearing into 256-dim features.
- Intensity branch taking 1D intensity input and applying MLP + learned attention to emphasize high-return points.
The position and dynamics outputs are concatenated into 512-dim features and then projected to 256-dim. A self-attention module refines per-point features, and mean pooling provides global context. Intensity features are pooled separately in an attention-weighted manner to another 256-dim vector. The final fused embedding is the concatenation of these two streams, yielding 512-dim features that combine geometric, dynamic, and reflection information.
The output head has two components. A depth confidence subnet uses an MLP to produce a depth confidence score. A 3D bounding box decoder expands the fused representation to 1024-dim and predicts seven box parameters,
Training uses a composite objective with Smooth L1 for box regression, binary cross-entropy for depth confidence, and a custom IoU loss that penalizes badly aligned boxes. The implementation is reported in PyTorch, with the AdamW optimizer and ReduceLROnPlateau scheduler.
5. Empirical behavior and ablation-style comparisons
The main empirical conclusion is that cooperative fusion outperforms single-vehicle baselines, and that the strongest gains appear under strict IoU criteria. At IoU $0.9$, middle fusion with intensity encoding on the Fused Ego-Rear setting achieves 0.2034 mAP, compared with 0.0224 mAP for Rear Only, corresponding to a 9.0833 ratio. At IoU $0.8$, the corresponding values are 0.4851 versus 0.1082, a ratio of 4.48 (Song et al., 22 Aug 2025).
The experiments also separate view effects from fusion effects. At IoU $0.9$, Side Only attains 0.2500 mAP, while F-Ego Side reaches 0.2985. This indicates that view geometry itself matters, but the reported trend is that Fused Ego (multi-view) always outperforms single-view. A plausible implication is that cooperative gains are conditioned by sensor placement and scene geometry rather than arising as a uniform additive constant across all views.
The middle-fusion versus late-fusion comparison is explicit. At IoU $0.9$, Late Fusion (ego rear) reports 0.0890, while Middle Fusion Fused View (with intensity) reports 0.2034. The stated trend is that middle fusion outperforms late fusion at nearly all IoU thresholds, especially strict (). This is consistent with the architectural intuition that feature-level aggregation preserves more cross-vehicle signal structure than post hoc box merging.
The role of intensity-aware encoding is singled out as critical. The paper states that intensity encoding is essential for detection, especially from the side view, and that without it, mAP can drop to zero for some views. It also notes that a competitor, RP-net, lacking this intensity-aware mechanism, fails on these datasets. That comparison does not establish a general theorem about radar intensity in all settings, but it does show that within this benchmark, intensity is not a peripheral attribute; it is a major determinant of usable detection performance.
6. Benchmark status, limitations, and research implications
CoVeRaP is presented as the first reproducible, large-scale benchmark for cooperative FMCW radar perception, with public code and data, and as an enabling resource for repeatable, objective assessment of V2V radar fusion algorithms (Song et al., 22 Aug 2025). The intended research uses include robust radar-only perception, especially under occlusion and adverse conditions where LiDAR/camera fail, as well as fusion strategy optimization and efficient architectures for real-time multi-vehicle cooperation.
Its practical framing is also explicit. The work emphasizes affordable hardware—including consumer camera, TI radar, no LiDAR, and <\$400 GPS—while still demonstrating substantial gains in 3-D detection quality. This suggests a deployment-oriented interpretation: cooperative radar perception is being positioned not only as a robustness mechanism but also as a cost-conscious alternative to heavier sensing stacks.
The limitations are narrowly defined rather than hidden. The current release focuses on parallel-lane vehicle scenarios, and the stated future extensions are complex maneuvers (T-intersection, merges) and additional object classes such as pedestrians, bikes, signs. This constrains the present benchmark’s coverage. A plausible implication is that CoVeRaP should be read as a foundational dataset for cooperative FMCW-radar perception rather than a complete representation of open-world urban interaction.
Taken together, CoVeRaP establishes a concrete benchmark and baseline for radar-only cooperative vehicular perception: synchronized multi-vehicle FMCW sensing, explicitly defined fusion paradigms, an intensity-aware PointNet-style detector, and evidence that middle fusion and shared radar viewpoints materially improve 3-D detection at strict localization thresholds.
