CoCam4D: Geometry-Aware Cooperative 4D Perception for Camera-Only Autonomous Driving

Soham Pahari; Sudip Das; Arindam Das; Ujjwal Bhattacharya

cs.ROOct 8, 2026

CoCam4D: Geometry-Aware Cooperative 4D Perception for Camera-Only Autonomous Driving

Authors: Soham Pahari, Sudip Das, Arindam Das, Ujjwal Bhattacharya

Organizations: University of Petroleum and Energy Studies, Dehradun · Indian Statistical Institute, Kolkata · anSWer India, Valeo Brain

Abstract

Autonomous vehicles often suffer from limited perception due to occlusions, blind spots, limited sensor range, and the complex nature of surrounding environments. Multi-agent collaborative perception (CP) addresses these challenges by allowing vehicles to share sensory information and reconstruct the scene cooperatively. However, camera-only perception remains fundamentally limited by the uncertainty of distance-dependent monocular depth estimation. We propose CoCam4D, a Bayesian framework for collaborative perception that explicitly models geometric uncertainty. It uses a VGGT-based feedforward network to generate 3D Gaussian scene representations with associated uncertainty estimates, enabling multiple vehicles or agents to efficiently combine their observations. By sharing compact Gaussian primitives, reliable observations from one agent can reduce the depth uncertainty of another without requiring LiDAR sensors. To support real-world deployment, we introduce Dynamic Object Primitives (DOPs), a compact 35-byte representation designed for efficient C-V2X communication. Extensive experiments show that our proposed method consistently outperforms recent vision-only methods, achieving improvements of 11.48% on OPV2V+ and 10.62% on DAIR-V2X-C, demonstrating the potential of geometrically grounded collaborative perception for LiDAR-free autonomous driving.

Figures & tables

Figure 1: CoCam4D overview. Each CAV tokenises its $F{\times}C$ surround-view frames and passes them through a shared-weight VGGT encoder to produce velocity-aware 3D Gaussians with per-pixel depth uncertainty in a single forward pass—no LiDAR or per-scene optimisation required. Dynamic objects are compressed into Dynamic Object Primitives (DOPs, ${\approx}35$ bytes/object) encoding state and analytically derived covariance $\mathbf{R}_{k}\!\propto\!z^{4}/f^{2}$ , then fused across vehicles via an information filter $\mathbf{Y}_{\text{fused}}{=}\sum_{j}\mathbf{R}_{j}^{-1}$ that automatically assigns weights to the confident lateral views over uncertain head-on observations, yielding a cooperative 4D world model that surpasses single-vehicle LiDAR at $N{=}5$ agents.

Figure 2: Single-agent perception outputs in CoCam4D. The figure shows surround-view inputs from six vehicle-mounted cameras along with their corresponding predicted depth and uncertainty heatmaps. The right column compares LiDAR observations with CoCam4D cooperative object predictions (N=5)

ID	Name	Sensor	Vehicles	Purpose
C1	Single-Vehicle Camera	4 $\times$ RGB	1	Lower bound
C2	Single-Vehicle LiDAR	64-beam	1	Target to match
C3	CoCam4D ( $N$ =2)	4 $\times$ RGB	2	Min. cooperation
C4	CoCam4D ( $N$ =5)	4 $\times$ RGB	5	Main result
C5	Cooperative LiDAR	64-beam	2	Upper bound

Table 1: Experimental configurations used in the evaluation.

Method	Modality	Coop.	$N$	[email protected] BEV		NDS	AMOTA	KB/s
Method	Modality	Coop.	$N$	0–50 m	50–100 m	NDS	AMOTA	KB/s
CenterPoint	LiDAR	✗	1	62.3	35.8	52.1	45.7	—
BEVFormer [ 27 ]	Camera	✗	1	38.4	14.7	33.2	24.6	—
Where2comm [ 15 ]	Camera	✓	2	44.2	18.9	37.8	29.4	840
V2X-ViT [ 47 ]	Camera	✓	2	46.8	21.3	40.1	31.7	920
CoCa3D [ 17 ]	Camera	✓	2	51.4	25.6	44.3	36.2	1240

Table 2: Simulation results on on OPV2V+ [ 18 ] , the extended version of the original dataset OPV2V. Bold entries exceed single-vehicle LiDAR (CenterPoint). CoCam4D (

N

=5) is the first cooperative camera method to surpass single-vehicle LiDAR on all reported metrics. CoCam4D (

N

=2) outperforms CoCa3D by 11.5% [email protected] at 0–50 m and 11.3% at 50–100 m while transmitting 21

\times

less bandwidth.

Method	DAIR-V2X-C [ 52 ]
Method	[email protected] $\uparrow$	[email protected] $\uparrow$
BEVFormer [ 27 ]	28.7	18.4
Where2comm [ 15 ]	34.2	21.8
CoCa3D [ 17 ]	38.6	24.3
CoCam4D (Ours)	42.7	27.0

Table 3: Comparative results on the real-world dataset DAIR-V2X-C [ 52 ] . CoCam4D generalizes to real sensor data without domain-specific fine-tuning, improving [email protected] by 10.6% over CoCa3D.

$N$	[email protected] $\uparrow$	[email protected] $\uparrow$	AMOTA $\uparrow$	KB/s
1 (camera)	35.2	18.4	26.8	—
1 (LiDAR)	62.3	38.6	45.7	—
2 (cameras)	48.6	28.3	36.4	42
3 (cameras)	54.8	33.6	41.2	58
5 (cameras)	63.7	40.2	47.8	87
10 (cameras)	68.4	43.8	51.4	156

Table 4: Cooperative density sweep on CARLA. Underlined row marks the critical density threshold at which CoCam4D first exceeds single-vehicle LiDAR (shown in the dashed reference row) on all three metrics. Performance scales consistently with

N

through diminishing returns above

N

=10.

Figure 3: Cooperative perception analysis on OPV2V+. Left: accuracy-bandwidth trade-off. Middle: performance scaling with cooperative density. Right: contribution of DOPs, Gaussian caching, and uncertainty-aware fusion to cooperative gain.

Variant	$\mathcal{H}_{\mathrm{unc}}$	Causal	Lifespan	IF	[email protected] $\uparrow$	AMOTA $\uparrow$	KB/s
Full CoCam4D	✓	✓	✓	✓	60.4	44.8	58
w/o $\mathcal{H}_{\mathrm{unc}}$	✗	✓	✓	✓	55.2	40.3	58
w/o Causal Attn.	✓	✗	✓	✓	56.8	37.6	58
w/o Lifespan	✓	✓	✗	✓	58.1	43.2	58
BEV fusion	✓	✓	✓	✗	53.7	39.8	186
Late fusion	✓	✓	✓	✗	49.2	35.6	12

Table 5: Ablation study on OPV2V+ [ 18 ] (

N

=3). IF = information filter; BEV fusion replaces the Gaussian product with learned BEV attention; Late fusion transmits bounding boxes only.

	Latency (ms)		Pose Noise (m STD)
Method	100	500	0.2	1.0
Where2comm [ 15 ]	44.3	32.8	46.2	38.7
CoCa3D [ 17 ]	49.8	38.4	50.6	43.2
CoCam4D (Ours)	56.4	53.7	58.9	55.4

Table 6: [email protected] BEV under communication latency and relative pose noise on OPV2V+ [ 18 ] (

N

=3). CoCam4D degrades by only 2.7 AP under 500 ms latency versus 11.4 for CoCa3D, owing to velocity-based compensation in equation 21 .

Method	Repr.	KB/s $\downarrow$	[email protected] $\uparrow$	AP/KB $\uparrow$
Where2comm [ 15 ]	BEV	840	44.2	0.053
V2X-ViT [ 47 ]	BEV	920	46.8	0.051
CoCa3D [ 17 ]	BEV	1240	51.4	0.041
VOGS-CP [ 4 ]	Gaussians	386	49.7	0.129
CoCam4D (static)	Gaussians	34	43.6	1.282
CoCam4D (DOP)	DOPs	7	52.8	7.543

Table 7: Communication efficiency at

N

=2. AP/KB measures detection accuracy per unit bandwidth. CoCam4D (full) achieves the highest AP while transmitting 21

\times

less than CoCa3D. Static Gaussians are transmitted once and amortized; the 58 KB/s figure reflects steady-state per-frame DOP cost plus the amortized static cache.

Explore similar work

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
C2E: Boosting Ego-Only 3D Object Detection via Multi-Teacher Contrastive Knowledge Distillation
Jul 2, 2026Jinlong Wang, Xun Huang, Qiming Xia +2Lidar-Based 3D Object Detection 3D Object Detection
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 · Zhongguancun Academy, China
ViCo3D: Empowering LiDAR-based Collaborative 3D Object Detection with Vision Foundation Models
Jul 14, 2026Haojie Ren, Songrui Luo, Lingfeng Wang +6Lidar-Based 3D Object Detection Vision Foundation Models
University of Science and Technology of China

Jun 22, 2026, cs.CVJ/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

Explore similar work

UECP: Uncertainty-Enhanced Collaborative Perception

C2E: Boosting Ego-Only 3D Object Detection via Multi-Teacher Contrastive Knowledge Distillation

ViCo3D: Empowering LiDAR-based Collaborative 3D Object Detection with Vision Foundation Models