acceptodds
Withdrawn from ICLR 2027

Temporal Assembly of Received Regional Features for Collaborative Perception under Coupled Delay and Loss

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

Collaborative perception extends an ego vehicle’s view using observations from neighboring vehicles, but depends on timely communication. When vehicle-tovehicle (V2V) communication degrades, delay and packet loss can jointly leave a neighbor’s current update incomplete or unavailable at a perception deadline, potentially reducing detection accuracy. To study this phenomenon, we introduce CoDeV2V, a newly collected long-sequence multi-vehicle LiDAR dataset. CoDeV2V models each V2V link with a condition—good, bad, or outage—that can persist across updates. The link condition jointly affects transmission delay and loss; an outage makes the affected neighbor’s update unavailable and may block several consecutive updates if it persists. Consequently, current neighbor information may be incomplete or unavailable at the perception deadline, while retained information becomes stale. Existing methods address aspects of communication degradation, but on CoDeV2V the evaluated methods still lose substantial detection accuracy compared with complete and timely neighbor information. To mitigate the remaining loss in detection accuracy, we propose AssemblyFusion. AssemblyFusion assembles regional features received from current and earlier observations, using how old the information is and where it is available to guide fusion with ego features. The ego vehicle also selects detections from each neighbor’s latest available observation, updates older detections using motion estimates, and combines the selected detections with predictions from fused features. Together, regional feature assembly uses the portions of neighbor updates that arrive by the deadline, while motion compensation updates historical detections when a neighbor’s current detections are unavailable, mitigating the impact of coupled delay and loss on perception. Under severe coupled communication degradation on the 5,400-frame CoDeV2V test split, AssemblyFusion achieves 53.20 ± 0.46 BEV [email protected] across three communication realizations, 6.70 points above CoAlign. Ablations show higher accuracy with motion-compensated historical detections than with discarded or uncorrected historical detections.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Collaborative perception, V2V communication, 3D object detection, coupled delay and packet loss, information freshness, decodability-ffrst admission

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ci27OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates cooperative perception under degraded V2V communication, where coupled latency and packet loss cause missing or outdated information from neighboring vehicles. The authors introduce CoDeV2V, a long-sequence multi-vehicle LiDAR dataset that models communication links using three persistent states: good, bad, and outage. To address these challenges, the authors propose AssemblyFusion, which combines available current and historical features based on their freshness and spatial coverage. It also compensates for outdated detections using motion estimation before merging them with current predictions. Under severe communication degradation, AssemblyFusion achieves 53.20 ± 0.46 BEV [email protected] on a 5,400-frame test set, outperforming CoAlign by 6.70 points. Ablation studies further demonstrate the benefits of motion-compensated historical detections.

Critical strengths
  • Modeling communication degradation in cooperative perception through persistent link states, coupled latency and packet loss, and perception deadlines is more realistic than assuming independent packet loss or fixed delays. This modeling approach offers some value for evaluation protocols.

  • Extensive Baseline Comparisons and Code Auditing: The paper evaluates 14 external methods under a unified evaluation protocol with consistent frames, ROIs, and object sets. The authors also identify an issue in CoBEVFlow's publicly released code, where predicted flow is overwritten by ground-truth flow. This finding provides valuable insights for the research community.

Critical weaknesses
  • Using the feature pathway alone achieves an [email protected] of only 40.26, underperforming both CoAlign (46.49) and Late Fusion (44.12). A substantial performance improvement to 53.20 is observed only after introducing the object-state pathway, which includes detection selection and motion compensation. This raises a concern as to whether the regional feature assembly mechanism emphasized in the paper's title is truly the primary source of the reported performance gains. The current ablation studies do not sufficiently isolate the independent contribution of this mechanism. Adding a comparative experiment that removes or replaces the regional feature assembly module while retaining the detection-level compensation pathway would help clarify its actual effectiveness.

  • AssemblyFusion divides each update into 12 transmission units, each carrying a complete copy of the detection and motion information. Consequently, receiving just one unit before the deadline is sufficient to recover the entire detection set, providing substantial robustness against packet loss. However, the paper does not clearly explain how the baselines map their messages to transmission units or whether they benefit from comparable redundancy. Although all methods share the same communication events, this does not necessarily guarantee comparable information availability. Therefore, the performance advantage in Table 1 may partly arise from the redundant transmission strategy rather than the proposed feature assembly mechanism.

It would be better to clarify the transmission-unit mapping and decoding requirements for each baseline and provide comparisons where strong baselines are equipped with equivalent detection redundancy and motion compensation.

Questions for authors
  • If CoAlign and CoSDH are equipped with the same redundant detection pathway and motion compensation mechanism, how would their performance compare with AssemblyFusion? Would this experiment help isolate the contribution of feature assembly?

  • How would the performance change if historical regions, age maps, or region splitting (i.e., transmitting entire features instead) were removed from the feature pathway?

Reviewer_nT7jOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces CoDeV2V, a simulated multi-vehicle LiDAR benchmark with persistent link conditions that couple delay and packet loss. AssemblyFusion combines received regional features with selected, motion-compensated historical detections. Under severe degradation, it achieves 53.20 BEV [email protected], approximately 6.7 points above CoAlign.

Critical strengths

The treatment of partial arrival and stale observations addresses a useful practical problem. Replicating compact detection messages across transmission units provides a simple way to retain usable information when regional features are incomplete.

The paired ablation in Table 2 is particularly informative. Static historical reuse improves [email protected] over fresh-only admission but reduces [email protected] from 43.64 to 37.26; motion compensation raises it to 53.20. This supports the importance of correcting historical geometry. Shared communication events, fixed perception weights in the ablation, and detector-specific ego references help interpret these results.

Critical weaknesses
  1. All quantitative evidence comes from the authors' own simulation, with no evaluation on an existing sequential dataset or real-collected data.

Every reported number is measured on CoDeV2V itself, collected by the authors in CARLA/OpenCDA, with link-model parameters chosen by the authors rather than fitted to measured network traces (acknowledged in §6). The stated motivation for a new dataset addresses the short scenes of OPV2V, but no experiment tests whether an existing sequential dataset with comparable degradation injected would reproduce the ranking; existing sequential asynchronous benchmarks (e.g., V2X-Sim) are neither cited nor used, and no real-collected dataset is evaluated at all. Scene distribution, range profile, and link statistics are therefore all co-designed with the proposed method.

A limited evaluation on an existing sequential dataset with comparable delay/loss injection, or on a real-collected dataset with degradation applied to its communications, would substantially strengthen the claim. If the authors consider this infeasible, they should state precisely which information existing datasets lack.

  1. The method depends on accurate localization and time synchronization, which the benchmark does not test.

AssemblyFusion's mechanisms assume exact geometry and timing throughout: received regional features are rotated and translated into the reference frame before assembly; age maps and decodability-first admission (Eq. 3) select on exact observation timestamps; and constant-velocity compensation (Eq. 4) extrapolates historical detections using accurate relative pose. The benchmark injects communication impairment only — no pose or clock error is modeled.

The paper's own ablation shows how much rides on geometric fidelity: without motion compensation, historical reuse collapses [email protected] from 43.64 to 37.26 (Table 2). Localization error and clock offset act on the same pathway — misestimated age and misprojected detections — yet are untested although unavoidable in deployment and could plausibly consume much of the reported margin. This dependence is an independent reason why evaluation on real-collected data, where such errors are present by construction, is necessary.

Questions for authors
  1. Have you evaluated on an existing sequential dataset with comparable delay/loss injection, or on a real-collected dataset? If not, which specific information do existing datasets lack?
  2. Are pose and clock errors modeled in CoDeV2V? How does performance degrade under realistic localization error and clock offset — could you provide a sensitivity analysis over these factors?
  3. Can you report delay-only and loss-only results for AssemblyFusion and one or two strong baselines under the current protocol?
  4. What are the receiver-side hardware, batch size, and per-module latency breakdown, and is computation within the stated deadline?
Reviewer_J8gHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies collaborative LiDAR perception when V2V delay and loss are coupled through a persistent good/bad/outage link state. It proposes AssemblyFusion, which max-pools whatever features arrive across recent timestamps and merges them with the newest received detection set, extrapolated to the current time. Evaluation using artificially introduced communication degradation shows better performance, though the evaluation protocol seems to favor the design. The method trails several baselines under ideal communication.

Critical strengths
  • Important problem: persistent, bursty V2V degradation rather than i.i.d. delay/loss.
  • Broad comparison with 14 methods under shared frames, targets, and communication events.
  • Clean ablations separating history admission from motion compensation, plus useful age-stratified analysis.
Critical weaknesses
  • Limited novelty; gains mainly come from standard late-fusion motion compensation.
  • Low-fidelity communication model that favors the proposed approach.
  • Overstated dataset and related-work framing.
Questions for authors
  1. There is very limited novelty against intuitive motion compensation at late fusion stage. The object-state branch is essentially late fusion with time compensation. "Decodability-first admission" reduces to using the latest fully received message. Since both components alone fall below CoAlign, please add the same compensated late-fusion branch to CoAlign, SCOPE, and CoSDH. Please also include a receiver-side tracker baseline.
  2. The channel model fidelity is questionable, lacks physical grounding. Real V2V communication channel has so many nuances. It is non-trivial to simulate it or model it. One cannot arbitrarily model the communication without citing prior research in this domain. Some of the modeling flaws include: Delay is independent of payload size. Payloads (1.85 MB, or 15 MB for CoAlign, at 10 Hz) far exceed DSRC/C-V2X capacity. Delays of up to 800 ms do not match V2V broadcast. Parameters are hand-set rather than fit to real-world V2V traces. State depends only on distance, with no LOS/NLOS, congestion, or cross-link correlation. Beyond 55 m the stationary distribution is ~39% outage, which guarantees that history dominates. In fact, C-V2X standard guarantees ~300m communication range.
  3. The protocol appears to favor the proposed design. One surviving unit suffices for AssemblyFusion. A payload needing all 12 units on time is current only ~7% of the time even on a good link. Seven baselines fall below No Fusion; a simple stale-drop fallback to ego-only is needed for a fair comparison.

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

Related papers

Discussion (0)

Sign in to comment.