acceptodds
Withdrawn from ICLR 2027

When Silence Misleads: Censored Observation-aware Verification for Collaborative Perception

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Object-level collaborative perception lets an ego agent recover objects it cannot see, yet every fusion rule in use reads the ego’s own silence as evidence: a remote proposal the ego did not also detect is discounted. We show that this silence is censored by observability. Among uncorroborated remote proposals, the fraction that correspond to real objects falls from 10.7% to 2.0% on real data and from 6.0% to 0.1% in simulation as the ego’s opportunity to verify the claim rises, while the confidence the sender reports is flat across the same axis — so a rule that reads scores alone cannot separate an unobservable hypothesis from a refuted one. The same structure defeats calibration: scores well calibrated on ego-visible objects are over-confident on ego-invisible ones, and neither message-observable context nor corroboration counting repairs it. We formulate missing corroboration as a censored observation: an absence is evidence only where the ego was eligible to observe. COVE is the minimal operator enforcing that admissibility condition — a remote hypothesis is attenuated only when the ego was physically eligible to verify it and did not. It needs no retraining or extra message payload. On six frozen scene-level holdouts spanning real and simulated data and a heterogeneous LiDAR/camera pairing, it improves aggregate AP70 over the score-mean late-fusion rule it is applied to on every holdout (+1.9 to +5.7 pp, every interval excluding zero) and ego-invisible AP70 by up to +20.8 pp, significantly on the four simulated holdouts; treating silence as a negative vote costs up to −29.1 pp, and uniform down-weighting buys aggregate accuracy at the cost of the ego-invisible subset. A controlled intervention — identical proposals, association and emitted geometry, with only the score aggregation replaced by evidential aggregation— improves aggregate AP70 in all twelve bootstrapped comparisons with every paired interval excluding zero, so the effect is not a property of one scoring rule. Across six more expressive extensions, none yields a consistent held-out improvement over the minimal operator.

Primary area:
applications to robotics, autonomy, planning
Keywords:
collaborative perception, uncertainty quantification, heterogeneous agents, distribution shift, test-time adaptation, autonomous driving, reliability calibration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YZHWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Authors propose COVE, a post-hoc scoring operator for object-level collaborative perception. The paper argues that the ego’s failure to detect a remote proposal should be treated as negative evidence only when the ego was able to observe that region. To this end, the ego computes a ray-based verification opportunity v from its own LiDAR sweep. The score of an uncorroborated remote hypothesis is multiplied by a constant α only when v exceeds a threshold, and is otherwise unchanged. The paper also reports that 31% of the in-range infrastructure-side labels on DAIR-V2X-C have no counterpart in the cooperative annotations, and evaluates on a union of the two label sets. Experiments on DAIR-V2X-C, OPV2V and an OPV2V LiDAR-camera pairing, with frozen proposals and association, show aggregate AP70 gains of 1.9 to 5.7 points on all six holdouts. The gain on ego-invisible objects is significant only on the four simulated holdouts.

Critical strengths

The finding about the DAIR-V2X-C cooperative annotations is useful to the community, independent of the method. Under the released labels, late fusion is far below the ego-only detector (41.0 vs 68.0 AP70), and part of this gap comes from ego-invisible objects being absent from the labels. The evaluation protocol is rigorous and honest. Proposals and association are frozen, holdouts are split by scene in both directions, baselines get the same selection budget, and intervals use a paired bootstrap. The unfavourable results are also reported openly.

Critical weaknesses

The method itself is thin and implements a known principle. The main claim about ego-invisible objects is neither cleanly isolated in the evaluation nor supported on real data. I have the following concerns:

  1. The contribution is insufficient. The authors admit that the observability-conditioned update is not new, and a very similar check already exists in collaborative perception. In “On Data Fabrication in Collaborative Vehicular Perception: Attacks and Countermeasures” (USENIX Security 2024), each vehicle labels space as free, occupied or unknown, and detected boxes are required to fall in occupied rather than free regions. Computing free space by ray casting is also standard, e.g. “What You See is What You Get: Exploiting Visibility for 3D Object Detection” (CVPR 2020). What I doubt is whether a one-line gate — multiplying by α when v ≥ τ and the box is uncorroborated — is sufficient as the main contribution of a method paper. Propositions 1 and 2 are direct restatements of Bayes’ rule and should not be presented as theoretical results.
  2. The proxy v is weak. v is computed from the very box whose correctness is being judged. When a remote box is displaced, v measures whether the ego can see the reported location, not the real object. In practice, COVE vetoes 3% to 15% of the remote-only clusters that match a real object on five holdouts, and 30% on DAIR-B. In addition, α is a constant, so a real object missed by the ego detector in a visible region is penalised in the same way as a ghost box. The guarantee in Proposition 2 therefore holds only if 1[v ≥ τ] is a good proxy for observability, which is not shown.
  3. The evaluation does not isolate the claimed effect. According to Section 5.1, detections matching ground truth outside the evaluated subset are ignored, but unmatched false positives are counted in every subset. Suppressing false positives in free space that the ego can see will therefore also raise the ego-invisible AP, regardless of whether ego-invisible objects are protected. Moreover, naive max-score fusion and score-mean fusion do not explicitly discount uncorroborated boxes, so compared with them COVE mainly adds a free-space false-positive filter. The comparison that actually tests the paper’s idea is COVE versus the τ = 0 variant, and it should be part of the main results.
  4. The real-data evidence is not convincing. The ego-invisible gain on DAIR is 0.4 and 2.4 points, and both intervals cross zero. The simpler in-box return count beats COVE on this subset (Table 8). Late fusion with COVE remains below the ego alone (46.9 vs 55.6 AP70). The six holdouts are not independent, since the two directions reuse the same scenes and DAIR has only 22 sequences. The union labels add infrastructure-side annotations from the same system on which the infrastructure detector is trained, which may favour the collaborator’s proposals. Finally, the parameter grid is only 2×2 and selects different values on every holdout, and the ego detector is always PointPillars. I believe another real-world dataset and more detectors are needed in order to be convincing.
Questions for authors
  1. Can you decompose the ego-invisible AP gain into retained ego-invisible true positives and false positives removed elsewhere, e.g. by reporting recall at a fixed precision?
  2. If the in-box return count matches or beats COVE on real data, what does ray reachability add? On a new dataset, how would you choose between them?
  3. Can you report the rates of {v ≥ τ, Y = 1} and {v < τ, Y = 1} to show how far 1[v ≥ τ] is from the latent observability?
  4. How sensitive is the 31% to the de-duplication thresholds (BEV IoU ≥ 0.1 or centre distance < 2.5 m) and to pose error? Has it been checked manually?
  5. Can the ego-invisible effect on real data be established with more sequences, another real dataset or more detector seeds?
Reviewer_7ZSaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Existing late-fusion methods often treat the absence of ego detections as negative evidence. However, a missing ego detection may simply result from limited visibility, and the sender's confidence score cannot distinguish between invisibility and genuine negative evidence. To address this issue, the authors propose COVE, which suppresses remote detections only when the ego vehicle can observe the corresponding region but fails to detect the object. COVE requires neither retraining nor additional communication overhead. Across six held-out evaluation sets, it improves overall AP70 by 1.9–5.7 percentage points and achieves gains of up to 20.8 points for ego-invisible objects. Conversely, treating the absence of ego detections as negative evidence can reduce performance by up to 29.1 points.

Critical strengths
  • The paper explicitly identifies a systematic bias in learning-based object-level cooperative detection: treating the absence of ego detections as negative evidence. It also quantifies the impact of this bias on objects that are invisible to the ego vehicle.

  • The paper uncovers a critical benchmark annotation issue in DAIR-V2X-C, where cooperative ground-truth labels miss 31% of the objects, causing valid detections from late fusion to be unfairly penalized. This finding explains the counterintuitive phenomenon reported in prior literature where late fusion underperforms no fusion, offering direct value for benchmark evaluation in the field.

Critical weaknesses
  • On DAIR-V2X-C, the gains on the ego-invisible subset are +0.4 and +2.4 pp, with both confidence intervals including zero. On the same real-data holdouts, the simpler in-box point-count gate significantly outperforms COVE on this subset (Appendix M). The authors attribute this to the count gate’s coupling with the stratum definition, but a similar coupling may also affect the ray-based v (see the next weakness). Meanwhile, on DAIR-A, uniform attenuation improves aggregate AP70 by +13.2 pp, compared with +1.9 pp for COVE. The authors argue this comes at the cost of the ego-invisible subset, but since COVE’s own gain on that subset is not significant on real data, its practical advantage over the simpler rule remains unclear in real-world deployment. Consequently, the strongest evidence for COVE comes mainly from simulated data.

  • The ego-invisible stratum is defined by the number of ego LiDAR returns inside the ground-truth box, while the verification opportunity v is computed from the same ego LiDAR sweep. The authors note that the two are not equivalent: remote-only objects can still have high v when ego rays traverse the region, and COVE vetoes 3–15% of real remote-only clusters (30% on DAIR-B). However, the degree of coupling is not quantified. Because the gating signal and the evaluation stratum come from the same sensor and sweep, it is unclear how much of the ego-invisible gain reflects a genuine improvement rather than agreement between two quantities derived from the same observation.

Questions for authors

Based on the weaknesses:

  • Could the authors report the number of ego-invisible objects in each holdout, evaluate whether combining uniform attenuation with COVE yields complementary benefits, and further validate the method on additional real-world datasets such as V2V4Real or V2X-Real, after checking whether their annotations exhibit the same ego-centric omissions identified in Section 3.1?

  • Could the authors report the joint distribution of v and the number of ego LiDAR returns inside ground-truth boxes for matched remote clusters across all three strata, or at least the fraction of real remote-only objects with v below the threshold τ? Additionally, could the authors re-evaluate the main results using a visibility definition independent of the ego LiDAR sweep, such as occlusion derived from ground-truth scene geometry in simulation, where such information is available?

Reviewer_aym2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a failure mode in object-level collaborative perception, where a missed object corresponds to an occluded object or a failed detection. While the motivation is clear and pratical, the proposed solution is relatively simple. The experiment results are extensive. However, the main idea of collaborative perception still suffers from important limitation in real-world deployment and robustness.

Critical strengths
  1. Clear motivation
  2. Extra annotations for collaboarive-perception evaluation improvement
  3. lightweight deployment
Critical weaknesses
  1. Limited novelty in proposed solutions
  2. Some unclear method description
Questions for authors
  1. Is COVE used only as a ranking heuristic, or is any probabilistic calibration included?
  2. In section 4.3, COVE’s parameters are selected using aggregate plus remote-only AP70 on labeled development scenes. Any specific rationale behind the two grids for the two parameters?
  3. Can we expect that a competitive late-fusion system that already preserves accurate ego geometry across more real scenes and detector families can also be benefited using COVE?

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

Related papers

Discussion (0)

Sign in to comment.