Sparse2comm: Towards Robust Cooperative 3D Object Detection

Lei Yang; Boqi Li; Chunmian Lin; Li Wang; Ziying Song; Shaoqing Xu; Heye Huang; Haibao Yu

cs.CVOct 6, 2026

Sparse2comm: Towards Robust Cooperative 3D Object Detection

Authors: Lei Yang, Boqi Li, Chunmian Lin, Li Wang, Ziying Song, Shaoqing Xu, Heye Huang, Haibao Yu, +1 more

Organizations: University of Hong Kong, China

Abstract

Cooperative perception improves autonomous driving by sharing complementary observations among vehicles and roadside infrastructure for 3D object detection. However, practical deployment is constrained by limited bandwidth and unreliable cooperation, where packet loss, transmission delay, and spatial misalignment jointly degrade the cooperative feature stream. Existing methods often reduce communication cost or compensate for one degradation type, leaving coupled disturbances insufficiently addressed. To address this problem, we propose Sparse2comm, a bandwidth-efficient and robust cooperative 3D object detection framework that treats unreliable cooperation as progressive restoration over degraded cooperative features. Sparse Feature Encoding first encodes communication as randomly mask-sampled foreground features transmitted by collaborating agents, from which the ego vehicle reconstructs dense semantic representations. This sparse-to-dense mechanism learns to infer missing object-centric content from sparse observations, enabling ultra-low-bandwidth communication and packet-loss recovery within the same representation. On the semantically restored features, Latency-Aware Alignment predicts motion flow to compensate delayed messages, and Self-Calibrating Fusion estimates residual spatial offsets in a self-supervised manner before adaptive cross-agent fusion. Sparse2comm therefore restores semantic completeness, temporal consistency, and spatial alignment in an ordered pipeline. Extensive experiments on DAIR-V2X, OpenV2V, and V2V4Real show that Sparse2comm maintains competitive clean accuracy and consistently improves robustness under individual and mixed real-world degradations. Compared with the selective feature communication baseline Where2comm, Sparse2comm improves mixed-setting AP@0.5/AP@0.7 by +20.15/+11.79, +12.66/+11.07, and +15.36/+12.61 on the three datasets, respectively.

Figures & tables

Fig. 1: Illustration of real-world challenges in cooperative perception. (a) Ideal aligned point cloud: ideal scenario with accurately synchronized and aligned data between ego and collaborating sensors. (b) Noisy misaligned point cloud: real-world distortions caused by (left) packet loss, (middle) transmission delay, and (right) localization errors, which jointly lead to incomplete or misaligned spatial representations. (c) Performance degradation: standard methods suffer significant accuracy drops under these noisy conditions, whereas our proposed Sparse2comm maintains robust cooperative perception under bandwidth, latency, and spatial uncertainties.

Fig. 2: The overall framework of Sparse2comm. The collaborator and ego encoders extract BEV features from LiDAR point clouds. Sparse Feature Encoding transmits masked sparse features and reconstructs complete semantics to minimize communication cost and resist packet loss. Latency-Aware Alignment predicts motion flow to temporally align delayed messages. Self-Calibrating Fusion self-supervisively estimates residual spatial offsets to refine cross-agent spatial alignment. The fused BEV features are finally decoded by a detection head for 3D object detection.

Fig. 3: The Motivation of Sparse Feature Encoding. Each object can be reconstructed from partial local features by leveraging a learned database of global priors, enabling complete semantic recovery under limited bandwidth.

Fig. 4: Sparse Feature Encoding. The collaborating agent performs Mask Sampling to transmit sparse object-foreground features under limited bandwidth, while the ego vehicle uses an encoder–decoder network for Feature Reconstruction to recover complete BEV features supervised by $\mathcal{L}_{\text{REC}}$ .

Fig. 5: Latency-aware Alignment. Flow Generator predicts motion flow $\mathcal{V}_{i}^{t_{i}}$ from temporal BEV features on the collaborator side, which is transmitted to the ego vehicle for Affine-based temporal alignment. The module is trained with flow loss $\mathcal{L}_{\text{FLOW}}$ using future frames to ensure robustness against transmission delay.

Fig. 6: Self-calibrating Fusion. Self-calibrating Alignment predicts pixel-wise offsets $\Delta_{(i\to v)}$ via residual and deformable convolutions to correct pose errors, guided by self-supervised loss $\mathcal{L}_{\text{SCA}}$ . Feature Fusion further integrates aligned features to generate the final BEV representation $Z_{v}^{t_{v}}$ .

Method	Category	DAIR-V2X		OpenV2V		V2V4Real		BW (%) $\downarrow$
Method	Category	Clean	Mixed	Clean	Mixed	Clean	Mixed	BW (%) $\downarrow$
Basic and classic cooperative perception
No Fusion	Single-agent	42.45 / 30.78	42.45 / 30.78	79.78 / 67.16	79.78 / 67.16	40.31 / 22.47	40.31 / 22.47	0
F-Cooper [ 39 ]	Full feature	49.83 / 34.21	29.32 / 22.74	88.15 / 80.97	68.54 / 55.62	66.08 / 35.21	37.41 / 17.42	100.0
DiscoNet [ 42 ]	Full feature	51.31 / 35.77	33.26 / 27.50	88.90 / 82.03	70.26 / 58.34	65.47 / 35.96	39.60 / 19.05	100.0
V2VNet [ 40 ]	Full feature	61.71 / 46.25	33.95 / 21.71	92.16 / 85.38	69.58 / 56.71	67.14 / 36.42	38.38 / 17.79	100.0

TABLE I: Main comparison with representative cooperative perception methods across datasets. Each performance cell reports AP@0.5 / AP@0.7. “Clean” denotes evaluation without artificial communication or pose perturbation, while “Mixed” denotes the unified robustness setting with packet loss, transmission delay, and spatial misalignment. BW is reported as the relative transmitted feature volume with respect to dense feature-level cooperation on each dataset, where dense feature transmission is normalized to 100%. Gray entries denote the single-agent No Fusion baseline, while black bold and underlined numbers indicate the best and second-best cooperative-method results in each column, respectively.

Method	Category	Clean	Packet Loss	Delay	Misalignment	Mixed	BW $\downarrow$ (Mbps)
Basic and classic cooperative perception
No Fusion	Single-agent	42.45 / 30.78	42.45 / 30.78	42.45 / 30.78	42.45 / 30.78	42.45 / 30.78	0
F-Cooper [ 39 ]	Full feature	49.83 / 34.21	43.17 / 29.63	42.58 / 29.37	38.52 / 27.13	29.32 / 22.74	258.1
DiscoNet [ 42 ]	Full feature	51.31 / 35.77	46.30 / 32.27	44.72 / 31.55	41.26 / 30.94	33.26 / 27.50	258.1
V2VNet [ 40 ]	Full feature	61.71 / 46.25	48.83 / 34.49	57.72 / 34.83	43.63 / 27.28	33.95 / 21.71	258.1
V2X-ViT [ 43 ]	Full feature	58.37 / 42.77	48.02 / 35.68	54.44 / 34.04	45.59 / 30.36	37.31 / 26.41	258.1

TABLE II: Targeted robustness comparison on DAIR-V2X. All entries report AP@0.5 / AP@0.7, and the method list follows the main comparison in Table I . Packet Loss, Transmission Delay, and Spatial Misalignment denote representative robustness settings evaluated under the same protocol for all methods. Mixed combines packet loss, transmission delay, and spatial misalignment to evaluate robustness under coupled real-world degradations. Since all results are evaluated on DAIR-V2X with the same feature shape, BW is reported as the transmitted bandwidth in Mbps. Gray entries denote the single-agent No Fusion baseline, while black bold and underlined numbers indicate the best and second-best cooperative-method results in each column, respectively.

Panel of Figure 9 — Fig. 7: Robustness evaluation under varying packet loss ratios, transmission delays, and spatial misalignments. The clean point and representative disturbance levels are aligned with the DAIR-V2X robustness comparison in Table II .

SFE	LAA	SCF	Ideal		Noisy		BW $\downarrow$
SFE	LAA	SCF	AP@0.5 $\uparrow$	AP@0.7 $\uparrow$	AP@0.5 $\uparrow$	AP@0.7 $\uparrow$	(Mbps)
			67.68	52.18	37.95	27.75	12.9
✓			66.31	51.12	49.91	36.52	2.6
	✓		68.12	52.87	41.67	29.90	12.9
		✓	68.76	53.24	48.53	33.32	12.9
✓	✓		67.93	52.54	52.82	35.78	2.6

TABLE III: Ablation study of the core components. SFE, LAA, and SCF denote Sparse Feature Encoding, Latency-Aware Alignment, and Self-Calibrating Fusion, respectively.

Training Strategy	Ideal		Noisy
Training Strategy	AP@0.5 $\uparrow$	AP@0.7 $\uparrow$	AP@0.5 $\uparrow$	AP@0.7 $\uparrow$
Joint Training	58.65	48.23	49.40	33.77
Multi-Stage Training	69.42	53.78	58.10	39.54

TABLE IV: Ablation study of the training strategy.

Sample Strategy	Ideal		Noisy
Sample Strategy	AP@0.5 $\uparrow$	AP@0.7 $\uparrow$	AP@0.5 $\uparrow$	AP@0.7 $\uparrow$
Uniform Sampling	67.25	52.30	52.97	36.10
Random Sampling	68.76	53.12	56.23	38.52
Object-aware Sampling	69.42	53.78	58.10	39.54

TABLE V: Ablation study on masking strategies in the Sparse Feature Encoding module.

Method	Data augmentation		Ideal		Noisy
Method	RM	PD	AP@0.5 $\uparrow$	AP@0.7 $\uparrow$	AP@0.5 $\uparrow$	AP@0.7 $\uparrow$
Baseline			67.68	52.18	37.95	27.75
	✓		66.41	49.92	40.66	26.57
		✓	64.56	47.58	42.96	29.98
	✓	✓	67.03	49.86	45.47	28.43
Sparse2comm	✓	✓	69.42	53.78	58.1	39.54

TABLE VI: Ablation study on the data augmentation. RM denotes random masking on collaborator BEV features, with mask ratio sampled from

r\sim\mathcal{U}(0,1.0)

(Uniform distribution). PD denotes pose disturbance applied to BEV features, including rotational noise

\Delta R\sim\mathcal{N}(0,3.0^{\circ})

and translational noise

\Delta T\sim\mathcal{N}(0,0.5)

m.(Gaussian distribution).

Method	Category	Computation			Communication			End-to-End Latency
		Params	FLOPs	Model Lat.	BW	Tx.@10M	Tx.@50M	E2E@10M	E2E@50M
		(M) $\downarrow$	(G) $\downarrow$	(ms) $\downarrow$	(Mbps) $\downarrow$	(ms) $\downarrow$	(ms) $\downarrow$	(ms) $\downarrow$	(ms) $\downarrow$
Basic and classic cooperative perception
No Fusion	Single-agent	6.6	118.4	37.8	0	0	0	37.8	37.8
F-Cooper [ 39 ]	Full feature	7.3	225.6	48.9	258.1	2581.0	516.2	2629.9	565.1
DiscoNet [ 42 ]	Full feature	6.7	238.2	50.8	258.1	2581.0	516.2	2631.8	567.0

TABLE VII: End-to-end efficiency comparison of the representative methods. All methods are profiled under the same DAIR-V2X input setting. Model latency denotes pure inference latency, while transmission time is computed from the transmitted bandwidth under conservative 10 Mbps and optimistic 50 Mbps V2X link rates. Gray entries denote the single-agent No Fusion baseline, and black bold numbers indicate the best cooperative-method result in each column.

SFE	LAA	SCF	Params (M)	FLOPs (G)	Latency (ms)
			11.4	218.9	51.2
✓			24.0	269.3	59.8
	✓		16.0	255.7	56.2
		✓	16.8	239.2	54.0
✓	✓		28.6	306.1	64.7
	✓	✓	21.4	276.0	59.0

TABLE VIII: Ablation study of Sparse2comm efficiency. SFE, LAA, and SCF denote Sparse Feature Encoding, Latency-Aware Alignment, and Self-calibrating Fusion. We report inference-stage parameters, FLOPs, and latency under the same profiling setting.

Setting	#Blocks		Ideal		Noisy
Setting	$N_{E}$	$N_{D}$	AP@0.5 $\uparrow$	AP@0.7 $\uparrow$	AP@0.5 $\uparrow$	AP@0.7 $\uparrow$
S1: Shallow	3	1	66.23	51.12	53.17	35.14
S2: Encoder-heavy	7	1	68.47	53.18	56.41	38.23
S3: Balanced (ours)	5	2	69.42	53.78	58.10	39.54
S4: Decoder-heavy	5	4	69.27	53.63	57.86	39.31
S5: Deep (optional)	7	4	69.14	53.46	57.22	38.87

TABLE IX: Ablation study on the architecture of Sparse Feature Encoding.

N_{E}

and

N_{D}

denote the number of ViT blocks in the encoder and decoder, respectively.

Fig. 8: Detection result visualization on DAIR-V2X. Representative examples are shown under clean and mixed noisy settings. Green and red boxes denote the ground truths and detection results, respectively. Under packet loss, transmission delay, and spatial misalignment, existing methods exhibit noticeable false and missed detections, while our Sparse2comm maintains accurate and consistent bounding boxes.

Fig. 9: Detection result visualization on V2V4Real. Additional real-world vehicle–vehicle examples are shown under clean and mixed noisy settings. Green and red boxes denote the ground truths and detection results, respectively. Sparse2comm preserves more spatially consistent detections when cooperative features are affected by packet loss, transmission delay, and spatial misalignment.

Fig. 10: Visualization of the Sparse Feature Encoding (SFE). We visualize the refined BEV features, the mask-sampled BEV features obtained by the Mask Sampling submodule, the reconstructed BEV features produced by the Feature Reconstruction submodule, along with their corresponding information-loss maps and loss histograms. The mask-sampled BEV features suffer from severe information degradation, while the reconstruction submodule effectively recovers the missing content, producing features close to the refined BEV and loss values near zero. This behavior remains consistent across packet-loss ratios from 10% to 100%, demonstrating SFE’s strong capability to recover missing information and reduce bandwidth through sparse-to-dense reconstruction.

Explore similar work

May 1, 2026cs.CV

Robust Fusion of Object-Level V2X for Learned 3D Object Detection

Perception for automated driving is largely based on onboard environmental sensors, such as cameras and radar, which are cost-effective but limited by line-of-sight and field-of-view constraints. These inherent limitations may cause onboard perception to fail under occlusions or poor visibility conditions. In parallel, cooperative awareness via vehicle-to-everything (V2X) communication is becoming increasingly available, enabling vehicles and infrastructure to share their own state as object-level information that complements onboard perception. In this work, we study how such V2X information can be integrated into 3D object detection and how robust the resulting system is to realistic V2X imperfections. Using the nuScenes dataset, we emulate object-level cooperative awareness messages from ground truth, injecting controlled noise and object dropout to mimic real-world conditions such as latency, localization errors, and low V2X penetration rates. We convert these messages into a dedicated bird's-eye view (BEV) input and fuse them into a BEVFusion-style detector. Our results demonstrate that while object-level cooperative information can substantially improve detection performance, achieving an NDS of 0.80 under favorable conditions, models trained on idealized data become fragile and over-reliant on V2X. Conversely, our proposed noise-aware training strategy, coupled with explicit confidence encoding, enhances robustness, maintaining performance gains even under severe noise and reduced V2X penetration.

Lukas Ostendorf, Lennart Reiher, Onn Haran +1

Institute for Automotive Engineering, RWTH Aachen University, 52074 Aachen, Germany · Qualcomm Technologies, Inc., Kfar-Netter, Israel

V2X 3D Object Detection Multi-Sensor Fusion Uncertainty-Aware Fusion Noise-Aware

Ask PDF

Jul 21, 2026cs.CV

CoGoal3D: Collaborative 3D Object Detection with 3D-Aware Fusion and Refinement

V2X collaborative object detection features overcoming the limitations of single-vehicle systems by aggregating environmental features from multiple collaborative agents. However, existing mainstream V2X perception methods mainly focus on 2D BEV object detection. When 3D detection task is concerned, inferior results are obtained because they ignore the 3D spatial misalignment caused by differing height and attitude among the collaborators. In this paper, we propose a novel collaborative 3D object detection framework called CoGoal3D, which extracts and refines the 3D feature gradually in a two-stage pipeline. In the first stage, a multiscale 3D-aware global fusion module is designed to mitigate the 3D spatial misalignment. The resulting proposals are then refined in the second stage with an auxiliary task of 3D point reconstruction. An effective multi-agent collaborative data augmentation strategy is further proposed to enrich the training data while minimizing information loss. Extensive experiments on public real-world datasets demonstrate that our CoGoal3D achieves new state-of-the-art performance, with 3D AP@0.7 improvements of 10.86%, 10.34%, and 10.18% on the DAIR-V2X, V2V4Real, and V2X-Real datasets, respectively. Code is available at https://github.com/Megalo-f/CoGoal3D.

Zhihao Yang, Zhiyu Xiang, Peng Xu +5

Zhejiang University, Hangzhou, China · Zhejiang Provincial Key Laboratory of Multi-Modal Communication Networks and Intelligent Information Processing, Hangzhou, China · China Jiliang University, Hangzhou, China

3D Object Detection V2X 3D Generation Global-Local Fusion Framework

Ask PDF

May 3, 2026cs.CV

AFFormer: Adaptive Feature Fusion Transformer for V2X Cooperative Perception under Channel Impairments

Accurate 3D object detection is essential for ensuring the safety of autonomous vehicles. Cooperative perception, which leverages vehicle-to-everything (V2X) communication to share perceptual data, enhances detection but is vulnerable to channel impairments, such as noise, fading, and interference. To strengthen the reliability of intelligent transportation systems, this work improves the robustness of V2X cooperative perception under communication conditions that reflect common channel impairments. This paper proposes an Adaptive Feature Fusion Transformer (AFFormer), a Transformer-based framework that mitigates the adverse effects of corrupted features by modeling temporal, inter-agent, and spatial correlations. AFFormer introduces three key modules: Multi-Agent and Temporal Aggregation for context-aware fusion across agents and over time, Dual Spatial Attention for efficient modeling of spatial dependencies, and Uncertainty-Guided Fusion for entropy-driven refinement of fused features. A teacher-student knowledge distillation strategy further enhances robustness by aligning fused features with reliable early-collaboration supervision. AFFormer is validated on the V2XSet and DAIR-V2X datasets, where it consistently outperforms existing methods under both ideal and impaired communication conditions, demonstrating improved robustness to communication-induced feature degradation while maintaining a competitive efficiency-accuracy trade-off.

Xi Zhou, Tao Huang, Qing-Long Han +2

College of Science and Engineering, James Cook University, Smithfield, QLD 4878, Australia · School of Engineering, Swinburne University of Technology, Melbourne, VIC 3122, Australia · Transport for NSW, Sydney, NSW, Australia

V2X Collaborative Perception Channel Modeling

Ask PDF