CoDS: Robust Collaborative Perception via Expert-driven Detection and BEV Segmentation

Jinlong Wang; Yuang Jia; Junhong Lin; Nannan Li; Wei Gao

cs.CVAug 14, 2026

CoDS: Robust Collaborative Perception via Expert-driven Detection and BEV Segmentation

Authors: Jinlong Wang, Yuang Jia, Junhong Lin, Nannan Li, Wei Gao

Organizations: Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School Peking University Shenzhen, China · School of Computer Science and Engineering Macau University of Science and Technology Macau, China

Abstract

Collaborative perception breaks through single-view limitations via multi-agent information exchange. However, multi-source noise such as pose errors and communication delays degrades fusion feature quality, constraining perception performance. Joint training of detection and BEV segmentation provides a natural remedy, where segmented road regions help constrain target distributions and detection bounding boxes help recover ambiguous segmentation boundaries. To this end, we propose a robust Collaborative perception framework with expert-driven Detection and bev Segmentation (CoDS). To address spatial inconsistency in fusion quality, we first introduce the Collaborative Reliability Map (CoRM) to explicitly quantify feature quality distribution. Based on CoRM, we design the Semantic Mixture-of-Experts (S-MoE) module to extract differentiated features for inconsistent feature demands. Finally, to further mitigate feature noise degradation, the Bidirectional Task Complementary Interaction (BTCI) refines task-aware features through bidirectional injection. Extensive experiments on OPV2V and V2V4Real datasets show that our CoDS surpasses existing baselines on both tasks and maintains stable robustness under multi-source noise. Code: https://github.com/JinlongW128/CoDS and https://openi.pcl.ac.cn/OpenAIDriving/CoDS.

Figures & tables

Figure 1. The motivation of our CoDS framework. None.

Figure 2. CoDS framework. Given $N$ CAV-encoded BEV features warped to ego coordinate via pose matrices, we get multi-agent features $\{\mathcal{F}_{i}^{w}\in\mathbb{R}^{H\times W\times C}\}_{i=1}^{N}$ . CoRM first analyzes pixel-wise consistency across these features, yielding a collaborative reliability map (R-Map) $\mathcal{R}\in(0,1]^{H\times W}$ . The fused BEV feature $\mathcal{F}^{coop}\in\mathbb{R}^{H\times W\times C}$ is concatenated with $\mathcal{R}$ and passed to S-MoE . S-MoE uses $\mathcal{R}$ as a routing prior to distribute features among three semantic experts, producing $\mathcal{F}_{det}$ and $\mathcal{F}_{seg}$ . Finally, BTCI uses $\mathcal{R}$ as a gating signal for bidirectional interaction between $\mathcal{F}_{det}$ and $\mathcal{F}_{seg}$ , yielding refined $\hat{\mathcal{F}}_{det}$ and $\hat{\mathcal{F}}_{seg}$ for detection and segmentation heads. None.

Figure 3. Visualization of R-Map under Gaussian pose noise $\eta$ $(2.0\,\text{m},\,2.0^{\circ})$ on OPV2V. Low-reliability regions ( blue ) coincide with areas of dense vehicle presence, where pose noise induces severe feature misalignment. None.

Figure 4. Semantic Mixture-of-Experts (S-MoE). The fused feature $\mathcal{F}^{coop}$ , concatenated with the reliability map $\mathcal{R}$ , is routed to the semantic experts via a reliability-guided gating mechanism, producing task-aware features $\mathcal{F}_{det}$ and $\mathcal{F}_{seg}$ . None.

Method	Detection			Segmentation
Method	AP $@$ 0.3	AP $@$ 0.5	AP $@$ 0.7	Dynamic	Road	Lane
AttFuse ( Xu et al., 2021 )	89.94	89.41	83.65	71.20	63.97	48.08
Where2comm ( Hu et al., 2022 )	90.19	89.20	80.55	70.19	63.43	47.36
CoBEVT ( Xu et al., 2022b )	92.09	91.09	80.95	73.97	65.53	50.84
CoAlign ( Lu et al., 2023 )	89.27	88.85	82.73	71.57	63.88	48.33
CoSDH ( Xu et al., 2025 )	91.06	90.54	82.48	72.54	62.35	46.31

Table 1. Results on the OPV2V. Best results are in bold .

Method	Public	AP $@$ 0.3	AP $@$ 0.5	AP $@$ 0.7	Dynamic
AttFuse ( Xu et al., 2021 )	ICRA 2022	66.45	59.23	31.43	54.29
Where2comm ( Hu et al., 2022 )	NeurIPS 2022	65.11	58.77	31.21	63.44
CoBEVT ( Xu et al., 2022b )	CoRL 2022	68.05	58.78	31.70	64.94
CoAlign ( Lu et al., 2023 )	ICRA 2023	69.45	62.02	31.81	57.58
CoSDH ( Xu et al., 2025 )	CVPR 2025	67.64	61.42	31.00	61.58
Ours (CoDS)	-	72.63	66.04	38.33	65.32

Table 2. Results on the V2V4Real. Best results are in bold .

Module	Test (AP $@$ 0.3/0.5/0.7)	Dynamic	Road	Lane
Baseline ( Xu et al., 2021 ) (only-det)	89.94 / 89.41 / 83.65	-	-	-
Baseline ( Xu et al., 2021 ) (only-seg)	-	71.20	63.97	48.08
+Seg Head	92.98 / 92.01 / 84.59	71.05	62.39	46.82
+S-MoE	93.32 / 92.62 / 86.03	73.73	65.13	50.44
+R-Map	94.22 / 93.56 / 85.68	75.82	67.44	53.11
+BTCI (Ours)	94.64 / 94.05 / 86.88	76.47	69.38	55.19

Table 3. Ablation study of each proposed module on OPV2V. “Baseline (only-det)” and “Baseline (only-seg)” denote single-task detection and segmentation models. Best in bold .

Figure 5. Robustness evaluation under increasing Gaussian localization noise on OPV2V. The x-axis denotes the injected pose noise level (translation in meters / rotation in degrees), ranging from $(0.0\,\text{m},\,0.0^{\circ})$ to $(0.6\,\text{m},\,0.6^{\circ})$ . The top row reports detection performance at AP $@$ 0.3/AP $@$ 0.5/AP $@$ 0.7, while the bottom row reports BEV segmentation IoU for Dynamic, Road, and Lane. None.

Method	0 ms			0–200 ms			0–300 ms			0–400 ms			0–500 ms
Method	AP $@$ 0.5	AP $@$ 0.7	Dynamic	AP $@$ 0.5	AP $@$ 0.7	Dynamic	AP $@$ 0.5	AP $@$ 0.7	Dynamic	AP $@$ 0.5	AP $@$ 0.7	Dynamic	AP $@$ 0.5	AP $@$ 0.7	Dynamic
AttFuse ( Xu et al., 2021 )	89.41	83.65	71.20	82.34	61.03	68.12	73.47	53.08	66.16	67.45	49.53	64.24	63.56	47.29	62.33
CoAlign ( Lu et al., 2023 )	88.85	82.73	71.57	81.07	58.89	68.76	72.02	51.89	66.58	65.32	47.51	65.08	61.83	45.59	63.47
CoSDH ( Xu et al., 2025 )	90.54	82.48	72.54	79.94	56.22	69.16	69.78	48.70	66.60	63.19	45.24	64.66	58.47	42.33	62.34
Ours	94.05	86.88	76.47	88.05	63.98	72.95	79.47	55.71	70.68	72.98	51.63	68.58	68.89	50.08	66.50

Table 4. Robustness evaluation under varying communication latency on OPV2V. We sample a random latency from a uniform distribution

\mathbf{U}[0,L]

, gradually increasing the maximum latency

L

from 200,ms to 500,ms. AP

@

0.5, AP

@

0.7, and Dynamic IoU are reported for each latency setting. Best results are in bold .

Module	Test (AP $@$ 0.3/0.5/0.7)	Dynamic	Road	Lane
Det-only	92.38 / 91.69 / 84.59	-	-	-
Seg-only	-	71.22	65.46	49.68
Joint Training	94.64 / 94.05 / 86.88	76.47	69.38	55.19

Table 5. Effect of joint multi-task training on OPV2V. “Det-only” and “Seg-only” denote models trained with a single task objective, while “Joint Training” refers to the full CoDS trained with both tasks simultaneously. Best in bold .

Method	AP $@$ 0.3/0.5/0.7	Dynamic	Road	Lane
BEVerse ( Zhang et al., 2022 )	92.66 / 91.65 / 82.75	69.32	63.61	44.78
CoBEVFusion ( Qiao and Zulkernine, 2023 )	93.26 / 92.59 / 83.55	73.32	59.83	45.13
ICA-V2X ( Yan et al., 2025 )	- / 92.10 / 84.50	-	64.70	53.70
Ours (CoDS)	94.64 / 94.05 / 86.88	76.47	69.38	55.19

Table 6. Comparison with single-agent and cooperative multi-task frameworks on OPV2V. BEVerse ( Zhang et al., 2022 ) is adapted to the collaborative scenario following the intermediate fusion strategy of AttFuse ( Xu et al., 2021 ) for fair comparison. Results for ICA-V2X ( Yan et al., 2025 ) are reproduced from the original paper, as no official code has been released. Best in bold .

Method	Params. $\downarrow$	FLOPs $\downarrow$	Mem. $\downarrow$	AP $@$ 0.5	Dynamic	Lane
AttFuse ( Xu et al., 2021 )	8.1M	97.3 G	616.3 MB	89.4	-	-
CoBEVT ( Xu et al., 2022b )	12.2M	190.8 G	573.0 MB	-	74.0	50.8
NCM	20.3M	381.2 G	1336.7 MB	89.4	74.0	50.8
Ours	11.0M	283.9 G	673.0 MB	94.1	76.5	55.2

Table 7. Comparison of model efficiency. Params., FLOPs, and Mem. denote the parameter count, floating-point operations, and average GPU memory consumption during inference.

Figure 6. Qualitative comparison of detection and BEV segmentation results. Left column: 3D detection results, where green and red boxes denote ground truth and predictions, respectively. Right column: BEV segmentation results, where orangered , royalblue , and white regions represent road, lane, and dynamic objects, respectively. None.

Bird's-Eye-View Segmentation

See all 11 papers

UECP: Uncertainty-Enhanced Collaborative Perception
Jun 22, 2026Kang Yang, Tianci Bu, Peng Wang +3Uncertainty Quantification Autonomous Driving Perception
Renmin University of China, Beijing, China · College of Systems Engineering, National University of Defense Technology, Changsha, China · Intelligent Shipping Center, China Waterborne Transport Research Institute, Beijing, China
Diff-KD: Diffusion-based Knowledge Distillation for Collaborative Perception under Corruptions
Apr 2, 2026Pengcheng Lyu, Chaokun Zhang, Gong Chen +2Cooperative Perception Autonomous Driving Perception
School of Future Technology, Tianjin University, Tianjin, China · School of Cybersecurity, Tianjin University, Tianjin, China · School of Computer Science and Technology, Tianjin University, Tianjin, China
BOLT: Online Lightweight Adaptation for Preparation-Free Heterogeneous Cooperative Perception
May 1, 2026Kang Yang, Tianci Bu, Peng Wang +2Cooperative Perception Adaptive Learning
Renmin University of China · National University of Defense Technology

Jun 22, 2026J/K move · Enter open · S save

UECP: Uncertainty-Enhanced Collaborative Perception

Kang Yang, Tianci Bu, Peng Wang, Deying Li+2

Renmin University of China, Beijing, China · College of Systems Engineering, National University of Defense Technology, Changsha, China · Intelligent Shipping Center, China Waterborne Transport Research Institute, Beijing, China

Uncertainty Quantification Autonomous Driving Perception Multi-Scale Feature Fusion Cooperative Perception

Open paper AskPDF arXiv

Collaborative perception serves as a pivotal solution to enhance the perception capability of individual agents in autonomous driving, where a core challenge lies in seeking reliable evidence to quantify and weight the contribution of each participating agent. Existing methods typically rely on a confidence map, which is co-trained with the detection head, but it is inherently correlated with the detection results and thus fails to provide unbiased physical evidence. Furthermore, how to deeply integrate evidence into the cooperative fusion process remains an open question. To address these issues, this paper first proposes an uncertainty map, a physically grounded and unambiguous metric for evaluating perception quality. This map is directly supervised by real-time sensor signals, i.e., LiDAR point density, ensuring decoupling from detection noise and thereby providing physical scenario-aware evidence for weighting agent contribution. Based on this map, we develop the Uncertainty-Enhanced Collaborative Perception (UECP) framework, centered on the Uncertainty-Aware Pyramid Fusion (UAPF) module. UAPF uses a coarse-to-fine strategy, with two key components: Uncertainty-Weighted Downsampling (UWD) for high-fidelity feature preservation, and Uncertainty-Guided Residual Fusion (UGRF) to reinforce ego features, suppressing noise and ensuring robust fusion. Extensive experiments on real-world datasets show UECP outperforms state-of-the-art methods in effectiveness and robustness by embedding the uncertainty map into fusion. Code will be publicly available.

Abstract

Bird's-Eye-View Segmentation

Explore similar work

UECP: Uncertainty-Enhanced Collaborative Perception

Diff-KD: Diffusion-based Knowledge Distillation for Collaborative Perception under Corruptions

BOLT: Online Lightweight Adaptation for Preparation-Free Heterogeneous Cooperative Perception