Cooperative 3D perception via Vehicle-to-Everything communication is a promising paradigm for enhancing autonomous driving, offering extended sensing horizons and occlusion resolution. However, the practical deployment of existing methods is hindered at long distances by two critical bottlenecks: the quadratic computational scaling of dense BEV representations and the fragility of feature association mechanisms under significant observation and alignment errors. To overcome these limitations, we introduce Long-SCOPE, a fully sparse framework designed for robust long-distance cooperative 3D perception. Our method features two novel components: a Geometry-guided Query Generation module to accurately detect small, distant objects, and a learnable Context-Aware Association module that robustly matches cooperative queries despite severe positional noise. Experiments on the V2X-Seq and Griffin datasets validate that Long-SCOPE achieves state-of-the-art performance, particularly in challenging 100-150 m long-range settings, while maintaining highly competitive computation and communication costs.
Figures & tables
Figure 1: Core challenges in long-range cooperation: (a) observation imprecision for distant targets, (b) quadratic scaling costs of dense features, and (c) positional mismatch from alignment errors.
Figure 2: Performance comparison on Griffin-25m dataset. The X-axis and Y-axis represent perception metrics, while the bubble size and color encode the transmission cost on a logarithmic scale.
Figure 3: Task Definition of Cooperative Perception. Co-Visible, Ego-Missed, and Ego-Invisible cases should be correctly handled.
Figure 4: Overview of Long-SCOPE framework, highlighting our novel components: the Geometry-guided Query Generation module for dynamically proposing object queries and the Context-Aware Association module for robust multi-agent query matching.
Figure 5: Depth derivation for high-vantage agents. We predict the stable global height z^Qglb (left histogram) instead of the dispersed depth z^Qcam (right histogram) and utilize geometric priors.
Method
0–100 m Overall
0–50 m
50–100 m
BPS ↓
FPS ↑
AP ↑
AMOTA ↑
AP ↑
AMOTA ↑
AP ↑
AMOTA ↑
Late Fusion Baseline
0.136
0.099
0.357
0.325
0.004
0.000
2.55\times103
6.13
Where2Comm [ 13 ] (NIPS 2022)
0.131
0.149
0.384
0.438
0.000
0.000
1.05\times106
7.34
V2X-ViT [ 38 ] (ECCV 2022)
0.201
0.188
0.494
0.527
0.011
0.000
3.21\times106
7.50
UniV2X [ 45 ] (AAAI 2025)
0.160
0.150
0.417
0.428
0.007
0.000
9.84×104
6.51
CoopTrack [ 48 ] (ICCV 2025)
0.279
0.268
0.560
0.617
0.054
0.000
2.00×105
6.03
Table 1: Performance comparison on Griffin-25m . Bold , underline , and italic represent the top three among feature-level fusion methods.
Method
0–150 m Overall
0–50 m
50–100 m
100–150 m
AP ↑
AMOTA ↑
AP ↑
AMOTA ↑
AP ↑
AMOTA ↑
AP ↑
AMOTA ↑
Late Fusion Baseline
0.224
0.271
0.318
0.375
0.178
0.185
0.050
0.032
Where2Comm [ 13 ] (NIPS 2022)
0.059
0.004
0.141
0.080
0.026
0.000
0.000
0.000
V2X-ViT [ 38 ] (ECCV 2022)
0.289
0.345
0.421
0.509
0.231
0.251
0.054
0.027
UniV2X [ 45 ] (AAAI 2025)
0.166
0.019
0.255
0.101
0.123
0.000
0.032
0.000
CoopTrack [ 48 ] (ICCV 2025)
0.232
0.171
0.327
0.255
0.192
0.088
0.059
0.006
Table 2: Performance comparison on V2X-Seq . Bold , underline , and italic represent the top three among feature-level fusion methods.
Agent Type
Query Method
0–100 m Overall
0–50 m
50–100 m
AP ↑
AMOTA ↑
AP ↑
AMOTA ↑
AP ↑
AMOTA ↑
Drone (High-Vantage)
Baseline (Static Queries)
0.371
0.393
0.639
0.766
0.092
0.010
+ Dynamic (Direct Depth)
0.390
0.403
0.647
0.745
0.116
0.032
+ Dynamic (Height-Derived)
0.449
0.443
0.694
0.748
0.182
0.099
Vehicle (Ground-Level)
Baseline (Static Queries)
0.413
0.252
0.604
0.384
0.047
0.019
+ Dynamic (Direct Depth)
0.428
0.256
0.610
0.407
0.052
0.000
Table 3: Detailed ablation study for the GQG module. We report the performance gains of single-agent models on Griffin-25m dataset.
Figure 6: Comparison of robustness to localization errors on Griffin-25m dataset. Long-SCOPE maintains stable tracking and detection under increasing translation and rotation noise compared to the baseline.
Figure 7: Qualitative comparison of query association performance under severe localization noise on Griffin-25m dataset. Matched query pairs share the same color. While the Hungarian algorithm (middle, red arrows) struggles with alignment errors by prioritizing geometric closeness, our CAA module (bottom) robustly resolves these ambiguities.
Fujian Key Laboratory of Urban Intelligent Sensing and Computing, Xiamen University, China · Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China