LR-V2X: Loss-Resilient Collaborative Perception under Low-Bandwidth Communication

Kang Yang; Tianci Bu; Peng Wang; Deying Li; Yongcai Wang

cs.CVOct 8, 2026

LR-V2X: Loss-Resilient Collaborative Perception under Low-Bandwidth Communication

Authors: Kang Yang, Tianci Bu, Peng Wang, Deying Li, Yongcai Wang

Organizations: Renmin University of China · The Hong Kong University of Science and Technology (HKUST) · Hebei Key Laboratory of Real-virtual Integrated Autonomous Systems (RIAS)

Abstract

Given the inherent unpredictability of packet loss in vehicular wireless communications, V2X collaborative perception can yield practical benefits only if agents can achieve reliable collaboration under lossy and low-bandwidth communication conditions. Existing dense BEV feature fusion methods depend on redundant BEV feature exchange, which is infeasible in low-bandwidth scenarios, while compact-communication methods aggressively compress messages but can hardly recover the missing feature content after packet loss. In this paper, we present LR-V2X, a loss-resilient, latent-space reconstruction framework that converts corrupted received latents (even under severe 90% packet loss) into a spatial prior and then reconstructs the missing BEV information from this informative prior and using ego context as condition. Notably, the model can be trained under complete communication conditions and can be directly applied to lossy conditions at test time, eliminating the need for training under numerous lossy conditions. Experiments on DAIR-V2X and V2XREAL show that LR-V2X delivers the strongest robustness under severe packet loss and preserves reliable collaboration as communication quality degrades. And it reduces communication overhead by $64\times$ compared to dense BEV feature fusion baselines. Code will be released at https://github.com/sidiangongyuan/LR-V2X.

Figures & tables

Figure 1: Motivation and overview. (a) Baseline fusion receives spatially incomplete feature maps when spatial packets are lost. (b) Codebook-based methods transmit compact codes, but missing indices leave spatial holes that degrade detection. (c) LR-V2X reconstructs complete BEV features from corrupted received latents via prior-initialized feature reconstruction, enabling graceful degradation under packet loss.

Panel of Figure 2 — Figure 2: Packet-loss-induced distribution shift in BEV features on DAIR-V2X. Panels show F-Cooper, CodeFilling, and LR-V2X from left to right. Shift Dist. measures the Mahalanobis distance between packet-loss and reference centroids, RMS Dist. measures the root-mean-square Mahalanobis distance of packet-loss features, and Outlier Rate counts packet-loss features outside the reference high-tail threshold. Lower values indicate a smaller departure from the no-packet-loss reference regime.

Figure 3: Framework of LR-V2X. Each agent encodes BEV features into a compact latent for transmission. The receiver reconstructs the agent’s BEV via the Latent Prior Decoder and a noise-conditioned reconstructor, then fuses with its own features for 3D detection.

Figure 4: Architecture of the Latent Prior Decoder (LPD) and the DiT denoiser used for conditional feature reconstruction.

Table 1: Performance comparison under Complete Communication (0% Packet Loss) reference conditions and the main 90% Packet Loss robustness setting.

Figure 5: Performance under varying packet loss rates on DAIR-V2X and V2XREAL. Across the sweep, LR-V2X remains consistently competitive as communication quality decreases and reaches the highest overall mAP at the 90% packet-loss endpoint.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Illustration of i.i.d. packet loss (left) vs. burst packet loss (right) on an $8\times 8$ BEV spatial grid. Blue cells are received; red cells are lost. Under i.i.d. loss, dropped packets are spatially scattered at random. Under burst loss, consecutive packets are dropped together, producing a contiguous missing region.

Models	Short			Middle			Long			mAP (%)			Comm
Models	0.3	0.5	0.7	0.3	0.5	0.7	0.3	0.5	0.7	0.3	0.5	0.7	Comm
F-Cooper	79.15	76.03	65.84	76.34	71.05	56.56	55.91	50.68	37.78	69.88	65.08	52.12	8MB
CodeFilling	78.02	76.29	66.81	71.36	68.42	55.72	50.26	47.19	36.63	65.36	62.62	51.48	128KB
LR-V2X	79.96	77.53	66.97	74.15	70.22	56.38	52.70	48.75	36.99	67.83	64.21	51.93	128KB

Appendix

Table 2: Burst-loss results on DAIR-V2X under 90% packet loss. F-Cooper is a dense BEV feature fusion baseline with 8 MB communication, while CodeFilling and LR-V2X both operate at 128 KB per transmitted collaborator frame.

Method class	Bandwidth $B$	Retained ( $B\times 10\%$ )	Rel. retained
Dense BEV feature fusion (e.g., F-Cooper)	8 MB	819.2 KB	$64\times$
Medium compression ( $\times$ 2)	2 MB	204.8 KB	$16\times$
Our default ( $\times$ 8)	128 KB	12.8 KB	$1\times$
Aggressive ( $\times$ 16)	32 KB	3.2 KB	$0.25\times$

Appendix

Table 3: Absolute information retained under 90% packet loss at different bandwidth levels.

Method	Dense Native (8 MB)				Budget-Matched (128 KB)
	[email protected]	[email protected]	[email protected]	FPS	[email protected]	[email protected]	[email protected]	FPS	$\Delta$ [email protected]
Attn	66.87	66.20	52.56	6.67	62.99	60.25	48.65	6.91	-5.95
F-Cooper	67.89	65.90	52.87	5.39	63.24	61.67	48.67	9.30	-4.23
CoBEVT	68.83	63.55	44.47	4.92	67.73	62.85	52.99	5.30	-0.70
V2X-VIT	68.78	64.33	50.97	4.70	64.69	60.53	49.21	5.32	-3.80
HEAL	64.43	61.68	50.50	7.24	60.81	58.08	47.25	10.00	-3.60

Appendix

Table 4: Budget-matched comparison on DAIR-V2X under 90% packet loss. Dense BEV feature fusion baselines are contrasted between their native dense BEV setting (left) and the protocol-matched 128 KB setting (right). LR-V2X is shown once as the native 128 KB reference.

Condition			mAP (%)
Ego feat.	Trans. mat.	Latent feat.	[email protected]	[email protected]	[email protected]
			67.28	63.71	52.01
✓			78.90	74.15	55.77
✓	✓		79.69	75.34	56.56
		✓	67.30	63.75	52.04
	✓	✓	68.30	64.74	53.04

Appendix

Table 5: Ablation of reconstructor conditioning inputs on DAIR-V2X under complete communication (no packet loss). All variants reconstruct from the same latent-derived prior; the table only removes ego-BEV, transformation, and direct latent-conditioning pathways inside the reconstructor.

Figure 7: Ablation of the Latent Prior Decoder (LPD) on DAIR-V2X. We compare the full model, prior-only (using the LPD output directly and skipping the denoising reconstructor), and zero-prior (replacing the LPD initialization with an all-zero BEV prior before denoising) under varying packet-loss rates.

Downsample	[email protected]	[email protected]	[email protected]	Comm
$\times$ 2	82.55	77.64	59.42	2MB
$\times$ 4	81.52	76.79	58.52	512KB
$\times$ 8	81.03	76.16	57.74	128KB
$\times$ 16	79.58	73.51	54.52	32KB

Appendix

Table 6: Effect of spatial downsampling ratio on DAIR-V2X.

Method	[email protected]	[email protected]	[email protected]
No-noise reconstructor	62.19	59.68	49.14
LR-V2X ( $t_{\text{inf}}=700$ )	69.35	66.90	53.37

Appendix

Table 7: Effect of noise-conditioned training on DAIR-V2X under 90% packet loss. The no-noise reconstructor keeps the same latent-derived prior and conditioning inputs but feeds the prior directly to the reconstructor at

t=0

, without forward-noise perturbation.

Figure 8: Sensitivity to the single-step inference timestep

t_{\text{inf}}

on DAIR-V2X under 90% packet loss. Solid curves show sweep results from a fixed checkpoint across different

t_{\text{inf}}

values, and the star marks the selected timestep

t_{\text{inf}}=700

. The trend is broadly stable across timesteps, indicating that the single-step denoiser is not highly sensitive to the exact inference timestep within a wide operating range.

Figure 9: Detection-guided qualitative comparison between LR-V2X and CodeFilling on a fixed DAIR-V2X sample. The same target region is highlighted across complete communication and 90% packet loss. Under severe loss, LR-V2X reconstructs a clearer object-centric BEV response and preserves the correct downstream detection, while CodeFilling misses the same target.

Figure 10: Visualization of detection results on DAIR-V2X and V2XREAL under 90% packet loss. Columns compare HEAL, CodeFilling, and LR-V2X on the same samples. The top two rows are DAIR-V2X examples, and the bottom two rows are V2XREAL examples. Red boxes denote predictions and green boxes denote ground-truth boxes.

Explore similar work

SemRD-V2X: Closure-Guided Communication with Bounded Inference for Cooperative Perception
Sep 28, 2026Hu Xu, Chun Li, Siyuan Qiu +2Communication-Efficient Collaborative Perception Rate-Distortion Optimization
Shanghai Jiao Tong University, Shanghai, China
AFFormer: Adaptive Feature Fusion Transformer for V2X Cooperative Perception under Channel Impairments
May 3, 2026Xi Zhou, Tao Huang, Qing-Long Han +23D Object Detection V2X Communication
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
Dense Coverage, Sparse Refinement: Byte-Constrained Cooperative Perception
Sep 24, 2026Melih Yazgan, Timon Müller, J. Marius ZöllnerCommunication-Efficient Collaborative Perception V2X Communication
FZI Research Center for Information Technology · Karlsruhe Institute of Technology

Sep 28, 2026, cs.AIJ/K move · Enter open · S save

SemRD-V2X: Closure-Guided Communication with Bounded Inference for Cooperative Perception

Hu Xu, Chun Li, Siyuan Qiu, Zeyan Li+1

Shanghai Jiao Tong University, Shanghai, China

Communication-Efficient Collaborative Perception Rate-Distortion Optimization V2X Communication Cooperative Perception

Open paper AskPDF arXiv

Vehicle-to-Everything (V2X) cooperative perception improves 3-D detection by sharing intermediate features, but dense remote features may repeat context that the ego agent can infer locally. Most communication-efficient designs optimize masks or codes empirically, leaving a more basic question open: which remote evidence is indispensable given the receiver's own observation? We introduce a closure-fidelity perspective on ego conditioned remote perception. Under a finite deductive abstraction and explicit conditions, its rate--distortion function decomposes over an irredundant core, and the exact zero-distortion rate becomes

P_A H(π_A)

. This analysis suggests a concrete design principle: transmit compact evidence and recover derivable context with bounded receiver-side inference. Guided by this principle, SemRD-V2X is an operational neural proxy that combines exact-budget BEV support selection, pointwise channel compression, and masked shared-weight reconstruction before standard fusion. Experiments on simulated V2XSet and real-world DAIR-V2X validate the resulting design. In a controlled five-run V2XSet comparison against a locally reproduced V2X-ViT-v1 baseline on one Tesla V100, SemRD-V2X reduces the analytical feature payload by

26.6\times

while improving [email protected]/[email protected] by 4.13/8.57 points, with 3.81% additional mean compute latency. These results position closure fidelity as both an analytical lens and an actionable design principle for communication-efficient cooperative perception.

Abstract

Explore similar work

SemRD-V2X: Closure-Guided Communication with Bounded Inference for Cooperative Perception

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

Dense Coverage, Sparse Refinement: Byte-Constrained Cooperative Perception