acceptodds
Under review as a conference paper at ICLR 2027

RC-GeoCP: Geometric Consensus for 4D Radar and Camera Collaborative Perception

Abstract

Collaborative perception (CP) extends sensing range through feature sharing, but most systems remain LiDAR-centric. Camera and 4D radar sensing combines dense semantics with lower-cost range–velocity measurements, yet exploiting their complementarity across agents remains difficult. Camera depth ambiguity produces spatially diffuse BEV evidence, while sparse transmission couples the choice of evidence to its subsequent contribution. Separate acquisition and aggregation objectives can assign conflicting source preferences to the retained evidence. We propose RC-GeoCP, connecting radar-grounded representation with persistent receiver–source attribution. Geometric Structure Rectification (GSR) uses radar-conditioned deformable sampling and gated calibration to construct grounded BEV features before communication. On these features, Uncertainty-Aware Communication (UAC) combines receiver demand with semantic confidence and radar support in a pre-delivery responsibility field that selects sparse support within fixed quotas. The Consensus-Driven Assembler (CDA) maps the same field into a bounded source mixture, carrying acquisition preferences into assembly. Persistent attribution coordinates spatial selection and source weighting across sparse transmission. Across two datasets and three aggregation backbones, RC-GeoCP consistently improves detection. Its sparse model retains 98.6%/96.6% of dense [email protected] at 10.88/4.95 MB per link. Code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
4D Radar

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_d1DxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RC-GeoCP, a radar-camera collaborative perception framework combining geometric feature rectification, uncertainty-aware communication, and consensus-driven feature aggregation. The method uses radar information to improve BEV representations and a shared responsibility score to guide both feature transmission and fusion. Experiments on V2X-Radar and V2X-R report improvements in detection accuracy and communication efficiency.

Critical strengths

The paper addresses a relevant problem in collaborative perception and provides relatively comprehensive experiments across different fusion backbones. The results show improvements over existing methods, and the ablation studies provide some evidence for the effectiveness of the proposed components.

Critical weaknesses
  1. Limited novelty and an overly complicated design. The proposed framework mainly combines existing techniques, including radar-camera fusion, confidence-based feature selection, Top-K communicatino, and weighted feature aggregation. The key contribution geometric consensus essentially use the same score for feature selection and fusion. I find it difficult to identify a substantial methodological innovation here. The overall framework appear unnecessarily complicated relative to the underlying technical contribution.

  2. The main performance gains do not convincingly justify the proposed collaborative mechanism. Much of the improvement comes from GSR, which is essentially a local radar-camera fusion module. E.g., on V2X-R, adding GSR already improves [email protected] from 61.30 to 64.99, while the full framework achieves only 65.09. Although the additional components improve communication efficiency, their contribution to detection accuracy is much less impressive. This raises questions about whether the proposed cooperative mechanism is actually the main reason for the reported improvements.

  3. Questionable practical relevance and insufficient system-level validation. The method require multiple rounds of communication between agents, yet the reported latency excludes network transmission and protocol overhead. The communication payload is still substantial, reaching 4.95 to 10.88 MB per link. Without realistic network-level evaluation, it's difficult to see how the proposed framework would operate efficiently in practical autonomous driving systems.

Questions for authors
  1. What is the fundamental methodological novelty beyond combining existing radar-camera fusion, feature selection, and weighted aggregation techniques? Why is the proposed shared scoring mechanism preferable to simpler alternatives?

  2. Given that GSR accounts for most of the accuracy improvement on V2X-R, what is the actual contribution of the collaborative communication mechanism beyond reducing feature transmission? Does its additional complexity justify the relatively limited detection gains?

  3. How practical is the proposed multi-stage communication protocol under realistic V2X bandwidth and latency constraints? Can the authors provide end-to-end latency measurements, including network transmission and protocol overhead?

Reviewer_5kDDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RC-GeoCP for radar-camera collaborative perception, combining radar-guided feature rectification, uncertainty-aware communication (UAC), and source-weighted aggregation. The main contribution appears to be sharing source responsibility between feature selection and aggregation.

Critical strengths
  1. Motivation: The paper is well motivated in combining radar and camera sensing for collaborative perception, particularly in addressing camera depth ambiguity and reducing communication overhead. The idea of using shared source responsibility for both feature selection and aggregation is reasonable and potentially useful for improving consistency across the two stages.

  2. Evaluation: The paper provides relatively comprehensive evaluations on two datasets and three aggregation backbones. The results show improvements in detection accuracy while reducing communication payload. The ablation studies (Tables A1, A4, A5) are useful, particularly the comparison between shared and independently learned source scores, which supports the proposed acquisition–assembly coordination.

Critical weaknesses
  1. Novelty: This is one of the main concerns. The proposed framework appears to combine several existing components:

1.a) GSR module uses radar-conditioned deformable sampling and gated feature fusion, closely related to existing radar-camera fusion methods such as RCBEVDet and DFA3D. 1.b) UAC module combines confidence, receiver demand, and radar reliability for Top-K selection, similar to existing communication-efficient methods such as CoSDH and SlimComm. 1.c) CDA module uses softmax-based source weighting.

The main distinction appears to be sharing the same responsibility scores between feature selection and aggregation. Although Table A4 provides experimental evidence supporting this design, it is unclear how much novelty this paper introduces beyond reusing learned source scores in two stages.

Furthermore, 1.d) the proposed geometric consensus primarily shares radar-supported source scores between feature selection and aggregation. How does it explicitly enforce geometric consistency?

1.e) 4D radar representation: The paper claims to use 4D radar but appears to treat radar primarily as generic BEV features, without distinguishing the physical roles of different measurement dimensions (range, azimuth, elevation, and Doppler/radial velocity) and its physical resolution in these dimensions. In particular, Doppler measurements are viewpoint-dependent (like radial velocity), but it is unclear how this information is exploited or handled across collaborating agents. The paper lacks analysis or ablation studies showing how each measurement dimension contributes to the proposed modules and detection performance.

  1. Evaluation and practical limitations: The communication efficiency is evaluated via feature payload rather than end-to-end communication performance. Is the reported payload of 4.95/10.88 MB per link acceptable for real-world V2X deployments, particularly with multiple agents? The reported inference latency appears to exclude local feature processing and network communication.
Questions for authors

Geometric rectification: Table A5 shows that gated radar calibration alone accounts for most of the improvement from GSR, while radar-conditioned sampling provides an additional 0.51/1.07 [email protected] on both datasets. Given that GSR is motivated by camera depth-induced spatial misalignment, can the authors provide explanation or insights on how radar-conditioned sampling improves the spatial alignment of camera features, beyond the gains from radar feature fusion?

Shared source responsibility: Table A4 shows improvements over independently learned CDA scores. Can the authors compare against a simpler design that uses the same confidence-based score for both Top-K selection and source weighting, without the proposed geometric responsibility formulation? This would help establish whether the proposed responsibility formulation itself is necessary.

Communication efficiency: Can the authors provide end-to-end latency and communication evaluations under realistic V2X bandwidth constraints? Given the reported payload of 4.95/10.88 MB per link, the practical communication cost appears significant.

Reviewer_76P4OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces RC-GeoCP, a camera + radar cooperative perception framework. The proposed framework consists of three key components: Geometric Structure Rectification (GSR), Uncertainty-Aware Communication (UAC), and Consensus-Driven Assembler (CDA). GSR rectify camera BEV features using radar deformable attention, while UAC and CDA selects feature to fuse based on uncertainty and consensus. Evaluation across two dataset and three backbones show that RC-GeoCP has gains over existing baselines. Ablation study highlights GSR contribution to the gains.

Critical strengths
  • camera + radar cooperative perception is relatively less explored, and possess practical impact
  • extensive evaluation on datasets/backgones, and delay pose noise, etc.
Critical weaknesses
  • limited novelty with a lack of comprehensive comparison to existing literature
  • evaluation does not support multi-agent real-world deployment, bandwidth, latency, onboard compute overhead are far from deployable
Questions for authors

The key component of the proposed framework is Geometric Structure Rectification (GSR), Uncertainty-Aware Communication (UAC), and Consensus-Driven Assembler (CDA). However, these components are discussed in literature broadly. In GSR, many intuition was treating radar as lidar and rectify BEV features similarly to BEVFusino of lidar and camera. In UAC and CDA, many bandwidth-aware framework has considered this, e.g. in where2comm, coopertrim, unisense, cosdh. The paper related work section stays too high level. Given such a crowded field in cooperative perception, the related work needs to be able to articulate in more nuanced details to distinguish the proposed framework with related work.

Perhaps due to the limitation of dataset, the discussion never covers what would happen with multiple agents (more than 2). This applies to both component design, especially UAC and CDA, as well as evaluation (everything is measured per link, unclear how results will change given more agents).

Although smaller than selected baselines, >10MB/link is way above what’s deployable. At 10 FPS, that’s over 800Mbps, even cellular has trouble to support that, not to mention C-V2X (1-2Mbps per link). Many recent work starts to look at how to retain cooperative gains within deployable capacity. As a community making cooperative perception towards a mature technology, we should pursue that together.

Also the data request and delivery mechanism introduces additional delays, which is not evaluated. It’s debatable whether that’s a good mechanism in practice.

Compute-wise, it’s more ideal if it can be benchmarked on what’s available on AV today, not server-level GPUs.

Introduction and figure 1 mentions a “quantitative table that reports complete algorithmic traffic” that never exists. The radar plot in figure 1 is also confusing, when AttFuse, CoAlign and HEAL is on the 6 corners combined with 2 different datasets, V2X-R and V2X-Radar, the different circles then compares camera, radar, bevfusion, Rc fusion. It’s unclear what to compare between a modality and a model.

Table 1 is listed in page 4 but was never explained in detail until page 9. And at page 9, there is still no explanation what 0, 0-30, 30-50, 50-100 means.

Table 6 shows on V2X-radar dataset, the performance has almost no sensitivity to policy at all. Is it saturated? This weakens the selection story. In Table A1, "Learned token" and "Full source control" are identical in all four columns. Also, the "Mean pooling" V2X-Radar entries (25.92/44.55) exactly match the sparse Full RC-GeoCP numbers.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.