acceptodds
Withdrawn from ICLR 2027

Perceive as You Receive: Collaborative Perception under Partial Reception

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Under tight perception deadlines and fluctuating V2X link conditions, an ego agent may need to produce a perception output before collaborative messages are fully received. We study this practical yet underexplored problem of collaborative perception under partial reception. Existing communication-efficient methods primarily optimize message size and accuracy under full reception, without explicitly prioritizing the early delivery of critical evidence or the effective use of partially received messages. We introduce CoDelta, a collaborative perception framework that couples transmission prioritization with ego-centric fusion under partial reception. The sender organizes compact BEV patch representations into a transmission sequence ordered by task relevance. The receiver combines received patches with ego features to generate local additive updates without reconstructing complete collaborator representations. It then aggregates these updates into the ego BEV and refines the resulting representation to restore spatial coherence. Extensive experiments on OPV2V, V2XSet, and DAIR-V2X demonstrate the effectiveness of CoDelta across varying reception ratios. On OPV2V, CoDelta outperforms PragComm by 14.75 percentage points in [email protected] at a 10% reception ratio. Code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Collaborative perception, Partial reception, V2X communication

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 4 reviews, above 27% of the others.
Reviewer_RNuUOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CoDelta, a new communication-efficient collaborative perception framework. The proposed method consists of sender-side priority-aware packetization and receiver-side delta fusion. The paper also considers partial reception (only a prefix is received) as an important challenge to be tackled in the field and provides a priority-aware ordering mechanism to enhance the performance under partial reception.

The proposed method's effectiveness is well validated. However, the way of dealing with baseline methods is not convincing.

Critical strengths
  1. It is valuable to discuss the partial reception scenario, which is a realistic challenge in real-world deployment.
  2. The compactness of the single-channel patch feature is impressive.
Critical weaknesses
  1. The partial reception scenario is the major motivation, the first claimed contribution, and the major experimental setup. However, it is not well analyzed, and the baseline methods are treated in a questionable way: As the author stated in Sec 4.1, compared methods' messages are arranged in spatial order. However, for many of the compared methods, spatial confidence scores are used to select the area to be transmitted. Using the confidence score to arrange the messages is more reasonable for a fair comparison. In fact, at least in Where2comm/ERMVP, the paper provides a systematic measurement of the performance-communication volume tradeoff. The performance-communication tradeoff study can be adapted to the partial reception scenario via engineering approaches.

  2. While the proposed method is well-motivated, the novelty is limited. In prior methods, feature fusion also uses a grid-wise 1*1 convolution. The incremental delta aggregation makes the previous F_fused-F_ego the predicted feature instead of the original F_fused, which could be treated as a residual connection. From my perspective, the novelty is limited to the patch-wise processing.

Questions for authors
  1. What's the major difference you consider when discussing "Packet Budget" and "Reception Ratio"?
  2. Could you compare the "priority score" and the confidence score used in previous work in some quantitative way?
Reviewer_74tFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CoDelta for collaborative 3D object detection under partial message reception. It prioritizes task-relevant BEV patches for transmission and uses the received patches to update ego features through Delta Fusion and contextual refinement. Experiments on OPV2V, V2XSet, and DAIR-V2X show improved detection accuracy under partial reception with low communication overhead. The authors also provide ablation studies and efficiency analyses.

Critical strengths
  1. The paper considers a relevant problem that is often overlooked in collaborative perception: the ego vehicle may need to make predictions before receiving the complete message. The motivation is clear, and evaluating performance at different reception ratios provides a useful perspective beyond the usual accuracy–communication trade-off.

  2. The proposed framework is reasonably well designed. Prioritizing BEV patches and directly updating ego features with the received information is a sensible approach to partial reception. The results are encouraging, with consistent improvements across three datasets, particularly at low reception ratios, while keeping the communication volume small.

Critical weaknesses
  1. The patch prioritization strategy is not sufficiently justified. The priority scores are based on foreground saliency in the collaborator's view, but salient regions are not necessarily the most useful ones for the ego vehicle. The paper does not examine whether higher-priority patches actually lead to larger detection gains, leaving the effectiveness of this design insufficiently explained.

  2. The partial-reception evaluation raises concerns about fairness and generalizability. CoDelta transmits patches in learned priority order, while the baselines are truncated in spatial order with missing regions zero-filled. This makes it difficult to determine how much of the improvement comes from the proposed fusion method. Moreover, the prefix-only assumption does not account for packet loss or out-of-order delivery in realistic V2X communication.

Questions for authors
  1. Does higher foreground saliency actually mean that a patch is more useful to the ego vehicle? Can the authors provide evidence that the priority scores reflect the detection gains from transmitting each patch?

  2. Would CoDelta still show similar improvements if the baselines also used priority-based transmission? How does the method perform when packets are lost or arrive out of order?

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

The paper studies collaborative perception under partial reception: the perception deadline expires before V2X messages fully arrive. CoDelta orders compact BEV patch packets by a distilled sender-side saliency and fuses received prefixes as ego-centric additive updates, without reconstructing collaborator feature maps. On three benchmarks it reports large gains at low reception ratios (e.g., +10.28 [email protected] over the strongest baseline at r=0.1 on OPV2V) with 8.25 KiB messages and the lowest fusion latency.

Critical strengths

The problem is genuinely underexplored and well motivated: neither complete-message compression nor delay/interruption lines use a partially received message. The mechanism evidence is coherent. Ordering is worth 2.7–11.4 [email protected] over random ordering under budget (Fig. 6); without Context-Guided Refinement, [email protected] is nearly flat in r (0.8287/0.8312/0.8310 at r=0.1/0.5/1, Table 6), so refinement is what converts reception into gains.

Critical weaknesses
  1. The protocol strips baselines of ordering, the paper's core mechanism

Baselines are truncated in spatial order with zero-fill, while CoDelta alone transmits in learned priority order. The paper's own ablation shows ordering is worth up to ~11 [email protected] at equal received count, so this asymmetry materially shapes Table 1. The byte-matched comparison with InfoCom is the right control but covers one baseline that is itself spatially truncated. Confidence/saliency-ordered truncation (Where2comm's maps already provide the signal), or random-order baselines, are needed to know how much of the low-r margin survives.

  1. Object-level collaboration, the natural prefix-decodable family, is not evaluated under partial reception

Late Collaboration — 7.031 KB of self-contained, confidence-sortable boxes, a message smaller than CoDelta's 8.25 KiB that degrades gracefully by losing trailing objects — is listed among compared methods (§4.1) but has no numbers at any r<1 in Table 1 or the appendix. Instance/query-feature-level methods are also absent. The premise that existing methods cannot exploit incomplete messages is thus untested on the family most robust to it.

  1. Ego-conditioned fusion may not inject evidence where the ego cannot see

The single-channel payload acts as a modulation signal on ego features (1×1 conv + tanh, Appendix A.2), so update content is bounded by ego feature support — precisely the occlusion regime that motivates collaboration. Evidence here is aggregate AP plus one qualitative scene (Fig. 3). Visibility-stratified results (objects outside ego sensor coverage vs. ego-visible) against ego-agnostic fusion (Standard/Late Collaboration) would show where the low-r gains come from.

Questions for authors
  1. How do key baselines perform under truncation orderings they can influence — their own confidence/saliency ordering, and random ordering — at equal reception ratios?
  2. Could you report Late Collaboration under the partial-reception protocol, and discuss or compare an instance/query-level method? If out of scope, why?
  3. Please provide visibility-stratified AP/recall (ego-blind vs. ego-visible objects) at r ∈ {0.1, 0.5, 1.0} for CoDelta vs. Standard/Late Collaboration.
  4. How is r sampled during training, and is one model shared across all reception ratios?
Reviewer_s9apOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents CoDelta, a collaborative perception framework for scenarios where only part of a collaborator’s message is available when perception must be performed. The sender compresses BEV patches and orders them according to predicted task relevance. The receiver uses the received patches to generate local residual updates to its own BEV representation, followed by contextual refinement. Experiments on OPV2V, V2XSet, and DAIR-V2X report improved detection performance under partial reception with compact communication.

Critical strengths
  1. A coherent mechanism for using incomplete collaborator information. Priority-based transmission and ego-conditioned residual updates fit the partial-reception setting well. Independently computable local updates provide a concrete way to incorporate received patches without requiring the entire selected message. This combination gives the framework practical methodological value within the evaluated setting, even though several individual ingredients build on established communication-efficient perception techniques.

  2. Encouraging accuracy–communication results across three datasets. The evaluation covers OPV2V, V2XSet, and DAIR-V2X, including real-world vehicle–infrastructure data. The reported gains are substantial at low reception ratios, and receiving the first quarter of the selected message already approaches full-reception accuracy across the three datasets. Together with the compact representation and reported processing efficiency, these results support the usefulness of the proposed framework and my inclination toward acceptance.

Critical weaknesses
  1. The scope of the contribution and the attribution of gains need clearer positioning. Importance-based spatial selection and compact feature transmission are established directions, and a shorter priority-ordered prefix also corresponds to selecting fewer top-ranked patches. The current “random ordering” ablation changes both selection and order, so it does not isolate their individual benefits. This limits how strongly the results establish a distinct ordering contribution beyond communication-budget adaptation, although it does not remove the reported value of the complete framework. The comparison with pipelines requiring complete collaborator-feature reconstruction should also identify which methods actually have that dependency.

  2. The reception abstraction and presentation limit interpretation of deployment claims. The experiments establish performance under controlled prefix truncation, but the mapping from logical patches to transport units, their independent availability, and the conditions supporting a contiguous prefix remain unclear. Benefits under practical perception deadlines therefore remain to be established. The paper would also be easier to assess with communication assumptions brought into the main text, explicit reception-ratio handling during training, consistent explanation of Figure 7’s denominator, and more readable figures. These are remaining scope and clarity limitations, not issues that the current results have resolved.

Overall assessment

Weak Accept. I lean toward acceptance because the coherent partial-reception framework and encouraging results across simulated and real-world data provide useful evidence within the evaluated setting. I retain reservations about the distinct contribution of ordering and the connection to practical transport and deadlines. These limit the claims that can be drawn, but on balance I consider the framework and reported accuracy–communication gains sufficient for a weak acceptance.

Questions for authors
  1. Clarification of scope and positioning: Which comparison pipelines require complete collaborator-feature reconstruction, and what capability does the proposed framework add beyond selecting fewer top-ranked patches under a smaller communication budget? Please distinguish the evidence for the full framework from the evidence specifically attributable to transmission order.

  2. Clarification of assumptions and reporting: Under what transport conditions would patches become independently available as a contiguous prefix? Please explain the mapping from logical patches to transport units, how reception ratios are handled during training, and why Figure 7 uses a different denominator from the main-text reception ratio. Please also distinguish measured reception-ratio performance from claims about practical perception deadlines.

  3. Further strengthening / future work: A controlled comparison that fixes the selected patch set while varying only its order, together with a transport-aware evaluation under a shared time budget, would help isolate ordering effects and assess deployment benefits. I view these as useful ways to strengthen and extend the evidence, rather than prerequisites for my current weak-accept recommendation.

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

Related papers

Discussion (0)

Sign in to comment.