acceptodds
Under review as a conference paper at ICLR 2027

Adaptation-Free Collaborative Perception with Unseen Agent Configurations

Abstract

Collaborative perception improves 3D object detection by enabling agents to share complementary observations, but most existing methods assume fixed or known collaborator encoder configurations, limiting deployment in practice. In this work, we consider an open-world setting in which auxiliary agents with unseen configurations may appear after deployment, such as different LiDAR beam counts or encoder architectures. To address this challenge, we propose ALF, a collaborative perception framework that enables adaptation-free collaboration with unseen agent configurations by lifting lightweight box-level messages into ego-compatible auxiliary features. ALF converts auxiliary box-level messages into pseudo-BEV maps and synthesizes ego-compatible latent features by combining object-centric cues with scene context from the ego feature. On V2X-Real, under a zero-shot evaluation across 64 case studies, ALF outperforms the strongest prior baseline by 35.91% in relative [email protected]. Controlled comparisons further show that, in the zero-shot setting, ALF is more accurate and robust than direct decision-level late fusion, while requiring only 120 bytes per agent per frame.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, Collaborative perception, 3D object detection

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_wjK1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ALF uses quantized detection boxes as a shared communication interface across heterogeneous agents. It converts received boxes into pseudo-BEV maps and combines them with ego features to synthesize features for intermediate fusion. This allows collaboration with unseen LiDAR specifications or encoder configurations without further adaptation.

Critical strengths

Reducing the adaptation cost of integrating new agents is practically relevant. The approach is straightforward and communication-efficient, requiring only 120 bytes of box data per agent per frame on V2X-Real. In the zero-shot evaluation, ALF achieves an average [email protected] of 0.311, compared with 0.229 for the strongest baseline. Comparisons with late fusion using the same boxes also support the benefit of feature synthesis.

Critical weaknesses
  1. The zero-shot comparison may be affected by how the baselines are implemented. Appendix B.5 uses an untrained, orthogonally initialized convolution to match feature channels when introducing a new encoder. This mapping may itself disrupt feature semantics, so the performance drop cannot be attributed solely to configuration changes. An identity-mapping control for configurations with matching channel dimensions would help isolate this effect. Table 8 also needs clarification: ALF’s results vary with AUX2 even though evaluation uses a fixed EGO+AUX1 pair and ALF undergoes no adaptation.

  2. The information lost through box compression and the limits of feature synthesis are insufficiently examined. Each message contains only planar position, length, width, yaw, and confidence, omitting class labels, vertical position, and height despite the multi-class 3D detection task. How reliably can these missing attributes be recovered when the object is occluded from the ego vehicle? Box dropout and noise experiments do not fully address this question. Results by class and occlusion level, together with comparisons against messages retaining class and height information, would be useful.

  3. Evidence for open-world deployment remains limited. Unseen configurations mainly involve PointPillar and SECOND variants and 64/128-beam LiDARs, while the ego configuration remains fixed. A shared message format ensures interface compatibility but does not remove differences in detector confidence or error distributions. Evaluation with detectors trained on different datasets or substantially different architectures would strengthen the generalization claim. The paper should also report inference time, memory usage, and how synthesis costs scale with the number of collaborators.

  4. Relevant prior work is missing. GT-Space: Enhancing Heterogeneous Collaborative Perception with Ground Truth Feature Space also addresses feature compatibility and the integration of new agents. It constructs a shared feature space from ground truth and trains adapters for new configurations. The authors should explain how ALF’s ego-specific feature space differs and compare detection accuracy, communication cost, and adaptation cost under matched data and training conditions.

Questions for authors

see Weakness.

8 comments
AuthorsOpenReview ↗

Response to Reviewer wjK1 (1/6)

W1: Correction to Table 8 and channel projection in zero-shot evaluation


Correction to Table 8 and 10

Thank you for flagging this discrepancy. Tables 8 and 10 evaluate Ego + AUX1. We mistakenly entered ALF scores indexed by AUX2. The correct scores correspond to the AUX1 configuration used at inference. When the receiving model and AUX1 are fixed, the ALF score is identical across rows that list different AUX2 configurations. In the revised manuscript, we will correct the ALF entries in Tables 8 and 10 to report the scores for the evaluated AUX1 configurations.


Clarification on channel projection

In zero-shot evaluation, an unseen AUX encoder can produce a bird’s-eye-view (BEV) feature map with a different channel count from what the pretrained receiving module expects. A channel mismatch prevents the module from processing the input directly. To run the pretrained baseline without retraining, we apply a fixed 1×1 convolution that maps the AUX feature vector at each BEV grid cell to the expected channel count. We bypass this projection when the channel counts match.

We considered channel truncation, zero-padding, and an unconstrained randomly initialized 1×1 convolution. Truncation discards channels by index, zero-padding adds empty channels, and an unconstrained random projection can scale feature directions unevenly. We chose orthogonal initialization as a fairer training-free control because it mixes input channels without selecting them by index and scales the retained directions uniformly. The projection weights remain fixed during zero-shot evaluation. The effect of this projection on AP in channel-mismatched cases remains unquantified.

For transfers between PP6 and PP8, all four feature-fusion baselines bypass the added projection in both directions. For transfers between SD2 and SD3, E2E and CodeFilling project the AUX features from 512 to 256 channels to match the PP4 Ego features. MPDA’s resizer and GenComm’s message encoder expect the channel count used during their training. In both transfer directions, the incoming features and the pretrained AUX modules use 512 channels, so MPDA and GenComm bypass the added projection. We will clarify these conditions in Appendix B.5.

AuthorsOpenReview ↗

Response to Reviewer wjK1 (2/6)

W2: Effects of adding class, vertical position, and height to box messages [1/2]


To evaluate the effect of retaining class, vertical position, and height in the box message, we compare 6-field ALF and 9-field ALF. We measure class-wise detection AP and recall under limited Ego visibility in this post, and evaluate the accuracy of the final box attributes in [2/2].


Experimental setup

The six-field message contains planar position (x, y), length, width, yaw, and confidence. The nine-field message adds the auxiliary detector's predicted class, vertical position z, and box height. We train the two versions of the Ego-compatible Feature Synthesizer (EFS) under matched conditions with the receiving model fixed. We evaluate both on V2X-Real using PP6, PP8, SD2, and SD3 as auxiliary configurations.


Detection performance by class

Each cell reports [email protected] / 0.7. mAP is the mean over the three classes. Bold marks the higher value for each threshold using the unrounded scores.

MetricALF, 6 fieldsALF, 9 fields
mAP0.480 / 0.3100.480 / 0.309
Vehicle AP0.735 / 0.5400.741 / 0.549
Pedestrian AP0.182 / 0.0150.181 / 0.015
Truck AP0.523 / 0.3740.517 / 0.362

At IoU 0.7, the nine-field message improves Vehicle AP from 0.540 to 0.549. Truck AP decreases from 0.374 to 0.362, both Pedestrian AP values round to 0.015, and mAP changes from 0.310 to 0.309.


Recall by class and Ego LiDAR visibility

We count Ego LiDAR returns inside each ground-truth 3D box. Zero denotes boxes with no returns. Sparse denotes boxes with at least one return and a count at or below a threshold that accounts for object projected area, class, and distance from Ego. We derive the thresholds from the lower quartile of positive point density on the validation split and hold them fixed for testing. We use these groups as a LiDAR-based proxy for Ego visibility.

Each cell reports class-aware [email protected] / 0.7. GT counts ground-truth objects once across the four AUX evaluations.

Ego visibilityClassGT6 fields9 fields
ZeroVehicle2,3600.741 / 0.5060.736 / 0.519
ZeroPedestrian5,7910.067 / 0.0070.067 / 0.007
ZeroTruck3790.117 / 0.0200.113 / 0.026
SparseVehicle4,2680.821 / 0.5730.820 / 0.578
SparsePedestrian10,0850.281 / 0.0730.282 / 0.072
SparseTruck1,4250.526 / 0.4410.521 / 0.440

Across the four auxiliary configurations, mean Vehicle [email protected] increases from 0.506 to 0.519 in Zero and from 0.573 to 0.578 in Sparse. Pedestrian and Truck recall changes vary across auxiliary configurations. We will add the class-wise AP and visibility-stratified recall results to the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer wjK1 (3/6)

W2: Effect of transmitting class, vertical position, and height [2/2]


Class, vertical position, and height accuracy in the final detections

We match final detections one-to-one to ground truth at IoU ≥ 0.5 using box geometry without requiring class agreement. Class accuracy is the fraction of matched detections with the correct class, computed separately for each model. For vertical position (zz) and height, we compute mean absolute error (MAE, in meters) on the same ground-truth objects matched by both models.

We average each metric equally across PP6, PP8, SD2, and SD3. Bold marks the higher class accuracy or lower MAE based on unrounded values.

Ego visibilityClassMessage formatClass accuracyz MAE (m)Height MAE (m)
ZeroVehicle6 fields0.9970.1520.177
ZeroVehicle9 fields0.9960.1520.172
ZeroPedestrian6 fields1.0000.2110.109
ZeroPedestrian9 fields1.0000.2020.108
ZeroTruck6 fields0.3650.3400.346
ZeroTruck9 fields0.3500.3710.328
SparseVehicle6 fields0.9610.1030.138
SparseVehicle9 fields0.9600.1020.136
SparsePedestrian6 fields1.0000.1040.090
SparsePedestrian9 fields1.0000.1030.090
SparseTruck6 fields0.7370.1330.143
SparseTruck9 fields0.7370.1320.146

With no Ego LiDAR returns, six-field ALF achieves class accuracy of 0.997 on matched vehicles and 0.365 on matched trucks. Both models correctly classify their matched Zero pedestrians, while Pedestrian [email protected] rounds to 0.007 for both formats. Adding the three fields lowers height MAE for all three Zero classes, while Truck zz MAE rises from 0.340 m to 0.371 m. We will add the per-class AP, visibility-stratified recall, and final-detection attribute analysis to the revised manuscript.


Results by AUX configuration

The table reports the individual results for PP6, PP8, SD2, and SD3 before averaging.

AUXEgo visibilityClassClass acc. (6 fields)Class acc. (9 fields)zz MAE (6 fields)zz MAE (9 fields)Height MAE (6 fields)Height MAE (9 fields)
PP6ZeroVehicle0.9970.9940.1400.1400.1630.157
PP6ZeroPedestrian1.0001.0000.2480.2340.1030.101
PP6ZeroTruck0.2120.1930.3510.3900.2830.273
PP6SparseVehicle0.9640.9610.1030.1020.1380.137
PP6SparsePedestrian1.0001.0000.1040.1040.0890.089
PP6SparseTruck0.7410.7430.1260.1230.1340.137
PP8ZeroVehicle0.9950.9970.1600.1590.1860.179
PP8ZeroPedestrian1.0001.0000.1780.1660.1030.098
PP8ZeroTruck0.3700.2890.3590.3750.3590.352
PP8SparseVehicle0.9570.9580.1050.1040.1420.140
PP8SparsePedestrian1.0001.0000.1030.1020.0890.089
PP8SparseTruck0.7420.7400.1310.1350.1500.155
SD2ZeroVehicle0.9980.9970.1520.1500.1730.170
SD2ZeroPedestrian1.0001.0000.2030.1960.1120.114
SD2ZeroTruck0.4870.5140.3450.3830.3470.304
SD2SparseVehicle0.9620.9610.1020.1010.1370.135
SD2SparsePedestrian1.0001.0000.1030.1020.0900.090
SD2SparseTruck0.7380.7400.1350.1350.1400.147
SD3ZeroVehicle0.9980.9950.1570.1600.1850.180
SD3ZeroPedestrian1.0001.0000.2130.2100.1180.119
SD3ZeroTruck0.3920.4050.3050.3370.3940.385
SD3SparseVehicle0.9590.9600.1020.1010.1360.134
SD3SparsePedestrian1.0001.0000.1060.1060.0910.091
SD3SparseTruck0.7290.7270.1390.1370.1480.143
Reviewer_ziD9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ALF, a collaborative perception framework that allows an ego agent to use detections from auxiliary agents with previously unseen sensor and encoder configurations without additional adaptation. Auxiliary agents transmit compact bounding-box messages, which ALF converts into a bird's-eye-view map and combines with ego features to produce representations usable by the frozen fusion and detection modules. The feature synthesizer is trained using detection supervision and alignment with teacher features produced by the ego encoder from auxiliary observations. On V2X-Real, evaluation across four fusion architectures and 64 zero-shot transfer cases shows a reported 35.91% relative improvement in aggregate [email protected] over the strongest evaluated prior baseline, using a box-message payload of 120 bytes per agent per frame.

Critical strengths
  1. ALF uses bounding boxes as a shared communication format while retaining feature-level interaction within the ego detector. This allows agents with different encoders to collaborate without aligning their internal representations or retraining an adapter for each new configuration. The combination of a fixed 120-byte box payload and adaptation-free inference addresses an important deployment problem: incorporating independently configured agents without repeatedly updating the receiving system.

  2. Experiments examine changes in encoder architecture, voxel resolution, and LiDAR beam count across four fusion backbones, providing evidence that the approach works beyond a single configuration. ALF also achieves the best [email protected] in 11 of 16 seen-configuration comparisons, suggesting that its flexibility does not require a consistent loss in detection accuracy (Table 3). The homogeneous ablations improve [email protected] from 0.295 without fusion to 0.333 with ALF and examine the contributions of the synthesis components, supporting the empirical value of the proposed design.

Critical weaknesses
  1. The experiments show that ALF handles unseen configurations better than the evaluated feature-based methods. What remains unclear is how much of this improvement comes from the learned feature synthesizer, rather than from communicating bounding boxes, which already provide a common format across encoders. Figure 6 includes late fusion only in a homogeneous setting, while the main unseen-configuration comparisons omit it and transfer feature-based baselines using a frozen, randomly initialized channel projection when channel alignment is needed (Appendix B.5). Comparing ALF with late fusion under the same unseen configurations and communication budget would help establish whether the synthesis step provides a substantial benefit over a simpler adaptation-free alternative.

  2. The central zero-shot claim depends on a clear separation between the auxiliary configurations used for training and those encountered during evaluation, but the paper does not fully specify this separation. The description of the single synthesizer training stage leaves its exact auxiliary training configuration unclear, and Appendix B.1 does not identify the data partition used. Table 7 also states that it evaluates AUX1, yet its ALF results change with AUX2 even when AUX1 and the fusion backbone remain fixed, which appears inconsistent with Appendix B.5. These details need clarification to establish which configurations were unseen and whether the reported comparisons follow the stated protocol.

Questions for authors
  1. How much does feature synthesis improve over directly fusing the same transmitted boxes when auxiliary configurations are unseen? Could you compare ALF with late fusion using identical detector checkpoints, quantization, and top-20 box messages, and report per-class AP and variation across repeated runs?

  2. Which auxiliary configurations and data splits were used to train each synthesizer, and was the same checkpoint reused across all transfer cases for each fusion backbone? Please clarify why the ALF entries in Table 7 vary with AUX2 when the caption specifies evaluation on AUX1 and Appendix B.5 states that results depend only on the configuration used at inference.

  3. Communication latency is one aspect of real-world deployment constraint. It would be great if the author could conduct synergy study with QuantV2X[1] (Fig.6) to see the further impact of the ALF in improving system-level efficiency.

[1] QuantV2X: A Fully Quantized Multi-Agent System for Cooperative Perception, ECCV 2026

5 comments
AuthorsOpenReview ↗

Response to Reviewer ziD9 (1/3)

W1/Q1 — Zero-shot comparison of ALF and direct late fusion


Clarification on channel projection

In zero-shot evaluation, an unseen AUX encoder can produce a bird’s-eye-view (BEV) feature map whose channel count differs from what the pretrained receiving module expects. A channel mismatch prevents the module from processing that input directly. To make the pretrained baseline runnable without retraining, we apply a fixed 1×1 convolution that maps the AUX feature vector at each BEV grid cell to the expected channel count. We bypass this projection when the channel counts already match.

We considered channel truncation, zero-padding, and an unconstrained randomly initialized 1×1 convolution. Truncation discards channels by index, zero-padding adds empty channels, and an unconstrained random projection can scale feature directions unevenly. We chose orthogonal initialization as a fairer training-free control because it mixes the input channels without selecting them by index and gives uniform scaling along the directions the projection retains. The projection weights stay fixed during zero-shot evaluation.

For PP6 and PP8, all four feature-fusion baselines bypass the projection. For SD2 and SD3, E2E and CodeFilling project the AUX features from 512 to 256 channels to match the PP4 Ego features. MPDA’s resizer and GenComm’s message encoder take AUX features at the channel count used to train those modules. In both transfer directions, the incoming features and the trained AUX modules use 512 channels, so MPDA and GenComm bypass the additional projection.


Experimental setup for the zero-shot late-fusion comparison

We fix the Ego encoder to PP4 and use AttnFuse. We evaluate four unseen AUX encoders (PP6, PP8, SD2, and SD3), all with 128-beam LiDAR. We keep the checkpoints fixed across three different evaluation seeds. For each seed, we compute AP separately for each AUX configuration and average the four results equally.

  • Standard LF uses all AUX boxes.
  • ALF-Box LF (9field) uses the Top-20 AUX boxes selected by confidence, following the same criterion as ALF. It retains the original FP32 box geometry, confidence, and predicted class.
  • ALF-Box LF (6field) uses the Top-20 AUX boxes selected by confidence and six INT8 values as ALF.
  • ALF without Aux disables AUX input to the same ALF receiver.

Mean detection performance

Each cell reports the mean across three evaluation seeds. mAP averages Vehicle, Pedestrian, and Truck. Bold marks the highest mean at each IoU.

MethodmAPVehicle APPedestrian APTruck AP
Standard LF0.426 / 0.2460.672 / 0.4490.160 / 0.0120.447 / 0.279
ALF-Box LF (9field)0.456 / 0.2690.717 / 0.4880.176 / 0.0140.474 / 0.304
ALF-Box LF (6field)0.447 / 0.2610.713 / 0.4760.152 / 0.0110.474 / 0.296
ALF without Aux0.456 / 0.2980.670 / 0.4830.184 / 0.0160.515 / 0.395
ALF0.480 / 0.3200.743 / 0.5460.178 / 0.0150.519 / 0.399

ALF achieves the highest mean mAP at both IoU thresholds, outperforming all three late-fusion variants. At IoU 0.7, ALF improves mAP by +0.0591 over ALF-Box LF (6field) with the same quantized Top-20 AUX boxes. ALF without Aux achieves the highest Pedestrian AP at both thresholds.

We will add the class-wise AP comparison across evaluation seeds to the revised manuscript.


Variability across evaluation seeds

Each cell reports the sample standard deviation across the three evaluation seeds.

MethodmAP SDVehicle AP SDPedestrian AP SDTruck AP SD
Standard LF0.000247 / 0.0003100.000240 / 0.0003120.000027 / 0.0000580.000950 / 0.000911
ALF-Box LF (9field)0.000225 / 0.0002570.000202 / 0.0003600.000056 / 0.0000950.000810 / 0.000822
ALF-Box LF (6field)0.000296 / 0.0001560.000538 / 0.0002670.000137 / 0.0000780.000732 / 0.000703
ALF without Aux0.000275 / 0.0003030.000403 / 0.0005650.000492 / 0.0001610.000222 / 0.000337
ALF0.000110 / 0.0004570.000102 / 0.0003620.000128 / 0.0000290.000320 / 0.001024

ALF’s mAP at IoU 0.7 was consistent across the three evaluation seeds, with a sample standard deviation of 0.000457.

AuthorsOpenReview ↗

Response to Reviewer ziD9 (2/3)

W2/Q2 — EFS training protocol and correction of Tables 8 and 10 results


1. EFS training and zero-shot evaluation

For both V2X-Real and V2XSet, we trained the Ego-compatible Feature Synthesizer (EFS), the standalone detectors, and all baselines on the respective dataset’s training split.

To construct the EFS training messages, we applied a pretrained standalone PP4 detector to raw AUX LiDAR observations from the training split and used its predicted boxes. To obtain the feature-alignment target, we transformed the same observations into Ego coordinates and passed them through the Ego encoder.

In the V2X-Real zero-shot experiments, separately trained standalone PP6, PP8, SD2, and SD3 detectors generated the AUX box messages. The EFS first received messages from these detector configurations during evaluation.

For each dataset, we trained one EFS for each fixed combination of Ego encoder, fusion module, and detection head. Within each dataset and fusion backbone, we reused the same EFS checkpoint across all AUX transfer cases without further training. We will specify the training splits, message-generation procedure, and checkpoint reuse in Appendix B of the revised manuscript.


2. Correction of the AUX1 evaluation results

We apologize for the confusion. As the reviewer pointed out, the ALF scores should correspond to the AUX1 configuration used at inference. We checked the submitted manuscript and confirmed that we had mistakenly entered scores indexed by AUX2 in the ALF columns of Tables 8 and 10. Thank you for flagging this discrepancy.

Tables 7 and 9 evaluate Ego+AUX2. Tables 8 and 10 evaluate Ego+AUX1. Under the protocol in Appendix B.5, when the receiver and evaluated AUX1 are fixed, the ALF score is the same regardless of which AUX2 is listed in the row. We will correct the ALF entries in Tables 8 and 10 in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer ziD9 (3/3)

Q3) System-Level Efficiency and Detection Accuracy with W8A8 Quantization


Experimental setup

Following the reviewer's suggestion, we use the computation-time notation from QuantV2X Fig. 6 to evaluate the latency and accuracy trade-off of model quantization in ALF. We fix the PP4 Ego encoder and AttnFuse on V2X-Real. The TensorRT export path used in our experiments did not support INT8 export for sparse 3D convolutions in the SD2 and SD3 auxiliary (AUX) detectors, so 12 such layers per detector ran in FP32. We therefore compare the three settings on PP6 and PP8, where we verified INT8 execution for all operations targeted by Full-model W8A8.

  • Original ALF is the FP32 model reference for this benchmark.
  • EFS-only W8A8 quantizes weights and activations to 8 bits in the ego-compatible feature synthesizer (EFS).
  • Full-model W8A8 applies the same precision to the targeted operations in the standalone Ego and AUX detectors and in the Ego-side fusion pipeline, including its feature encoder, EFS, fusion module, and final detection head.

All three settings use the six-field INT8 Top-20 box message format. We benchmark each setting three times per AUX on the same RTX 4090 with identical input frames. TlocalT_{\mathrm{local}} covers Ego and AUX local detection and Ego feature extraction for fusion. TfusT_{\mathrm{fus}} covers received-message processing, feature synthesis, fusion, and final detection, including post-processing. TcompT_{\mathrm{comp}} is the sum of the two stages. We run Original ALF in PyTorch with FP32 model precision and execute the quantized parts of the W8A8 variants using TensorRT. The measured latency therefore reflects both model precision and the execution backend.


Computation time and detection accuracy averaged over PP6 and PP8

We average PP6 and PP8 equally within each repetition and report latency as mean ± sample standard deviation across the three repetitions.

MethodTlocalT_{\mathrm{local}} (ms)TfusT_{\mathrm{fus}} (ms)TcompT_{\mathrm{comp}} (ms)[email protected][email protected]
Original ALF36.084 ± 0.83717.649 ± 0.25153.733 ± 1.0870.4740.320
EFS-only W8A835.958 ± 0.6009.241 ± 0.15445.199 ± 0.7520.4690.304
Full-model W8A834.136 ± 0.1489.599 ± 0.06843.734 ± 0.1980.3960.239

Compared with Original ALF, EFS-only W8A8 reduces mean TcompT_{\mathrm{comp}} by 8.534 ms, with decreases of 0.0057 in [email protected] and 0.0158 in [email protected]. Full-model W8A8 saves a further 1.465 ms relative to EFS-only W8A8, with additional mAP decreases of 0.0722 and 0.0648, respectively. EFS itself takes about 1.26 ms in both quantized settings. The full fusion stage also includes message processing, the fusion module, and the detection head, and it takes longer with Full-model W8A8 in all six paired measurements. On average, Full-model W8A8 reduces TlocalT_{\mathrm{local}} by 1.822 ms while increasing TfusT_{\mathrm{fus}} by 0.358 ms relative to EFS-only W8A8. Across PP6 and PP8, EFS-only W8A8 delivers most of the measured computation-time reduction and retains higher detection accuracy than Full-model W8A8.

We will add the quantization coverage, timing variability, and computation–accuracy results for both AUX configurations to the revised manuscript.


Results by AUX configuration

The table below reports results separately for PP6 and PP8.

AUXMethodTlocalT_{\mathrm{local}} (ms)TfusT_{\mathrm{fus}} (ms)TcompT_{\mathrm{comp}} (ms)[email protected][email protected]
PP6Original ALF37.841 ± 1.99217.837 ± 0.56955.678 ± 2.5600.4770.325
PP6EFS-only W8A836.804 ± 0.8189.159 ± 0.21745.964 ± 1.0340.4690.306
PP6Full-model W8A834.557 ± 0.2599.521 ± 0.09444.078 ± 0.3500.4000.243
PP8Original ALF34.327 ± 0.37317.460 ± 0.11151.787 ± 0.4810.4710.315
PP8EFS-only W8A835.112 ± 0.3989.322 ± 0.13644.434 ± 0.5160.4680.302
PP8Full-model W8A833.714 ± 0.1639.676 ± 0.04143.391 ± 0.1550.3920.235
Reviewer_mbs5OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses configuration-agnostic heterogeneity in collaborative 3D detection: agents with unseen LiDAR/encoder configurations appear at deployment and must collaborate without retraining. ALF sends only quantized box-level messages (120 B/agent/frame), rasterizes them into a pseudo-BEV map, and lifts them into the ego latent space with a small synthesizer trained once against ego-domain teachers, feeding a frozen fusion and detection stack. On V2X-Real it wins all 48 zero-shot encoder-heterogeneity cases and 14 of 16 LiDAR cases against four adaptation/reconstruction/generation baselines.

Critical strengths

The setting is well motivated and cleanly formulated: a message space independent of agent configuration keeps the heterogeneity module inside its training support, unlike latent-feature interfaces. The evaluation is systematic (64 zero-shot cases, three protocols, four fusion backbones), and the same-boxes controlled comparison against two late-fusion variants isolates the value of feature-level lifting, with threshold insensitivity and graceful degradation under corruption, false positives, pose noise, and latency. The design is frugal and consistent with the supplementary code.

Critical weaknesses
  1. The net gain over no collaboration is modest and unreported within the zero-shot protocol. The only no-collaboration anchor is in the homogeneous ablation ([email protected] 0.295, Fig. 10); the zero-shot average of 0.311 implies roughly +0.016, whereas the headline 35.91% is relative to baselines that collapse here. Without an ego-only row across the 64 cases, the practical size of the benefit — and whether it is ever negative — cannot be read from the paper.

  2. The decision-level comparison may understate late fusion. Both baselines are NMS-style merges; no weighted-box-fusion or learned box-association variant is tested, and ALF-Box LF numbers appear only in plots. Standard LF at its best threshold (0.292) merely matches the no-collaboration anchor, so it shows how naive merging fails, not what well-designed box-level fusion achieves. The robustness advantages (threshold insensitivity, false-positive rejection) are mechanism-level and likely survive, but the abstract's accuracy margin is established only against weak variants.

Questions for authors
  1. Report ego-only results for all 64 zero-shot cases, with per-case gains; does collaboration ever hurt?
  2. Add a stronger training-free decision-level baseline (e.g., weighted box fusion) to Figs. 4–5; tabulate the ALF-Box LF numbers.
  3. Why not evaluate camera-based collaboration? Visual domain gaps are typically larger than LiDAR's, stressing the configuration-agnostic claim harder, and V2X-Real provides cameras. Does the ego-domain teacher cover camera auxiliaries?
  4. Is EFS trained per fusion backbone? This bounds the "trained once" claim.
  5. Concurrent training-free adapters such as DiffAlign (online-training-free diffusion feature alignment) also target unseen-agent heterogeneity, while transmitting latent features. How does ALF position against this line — comparison or trade-off discussion (bandwidth vs. feature-level information)?
6 comments
AuthorsOpenReview ↗

Response to Reviewer mbs5 (1/4)

W1/Q1: Gains over no collaboration


Experimental setup

We compare No Fusion (Ego-only), ALF without Aux (the same ALF receiver with AUX disabled), and ALF. We report [email protected], averaging Vehicle, Pedestrian, and Truck.


Overall results

The following means follow the 64 reporting cases. Bold marks the highest displayed value.

MethodMean [email protected]
No Fusion0.293
ALF without Aux0.283
ALF0.311

Compared with ALF without Aux, enabling collaboration improves [email protected] in 61/64 cases and reduces it in 3/64 cases, with a mean gain of +0.0278 across all 64 cases. This shows that AUX input benefits the ALF receiver in most evaluated configurations.

Compared with No Fusion (Ego-only), ALF improves performance in 46/64 cases, with a mean gain of +0.0184. The 18 decreases comprise all 16 CoBEVT cases and two V2X-ViT cases. AttnFuse and ParCon improve over Ego-only in all 16 cases each, while V2X-ViT improves in 14/16 cases. Thus, improvements over Ego-only are observed across most evaluated fusion modules, while decreases are concentrated in CoBEVT. The largest decrease is 0.0587 for CoBEVT with PP8, 64-beam AUX.

We will add these comparisons and the complete 64-case results to the revised manuscript to clarify when collaboration improves performance and when it reduces accuracy.


Results for all 64 cases

Entries are [email protected].

Fusion modelAUX1→AUX2AUX2 beamsNo FusionALF without AuxALF
AttnFusePP6→PP81280.2930.2980.317
AttnFusePP6→SD21280.2930.2980.327
AttnFusePP6→SD31280.2930.2980.316
AttnFusePP8→PP61280.2930.2980.327
AttnFusePP8→SD21280.2930.2980.327
AttnFusePP8→SD31280.2930.2980.316
AttnFuseSD2→PP61280.2930.2980.327
AttnFuseSD2→PP81280.2930.2980.317
AttnFuseSD2→SD31280.2930.2980.316
AttnFuseSD3→PP61280.2930.2980.327
AttnFuseSD3→PP81280.2930.2980.317
AttnFuseSD3→SD21280.2930.2980.327
CoBEVTPP6→PP81280.2930.2350.261
CoBEVTPP6→SD21280.2930.2350.275
CoBEVTPP6→SD31280.2930.2350.268
CoBEVTPP8→PP61280.2930.2350.266
CoBEVTPP8→SD21280.2930.2350.275
CoBEVTPP8→SD31280.2930.2350.268
CoBEVTSD2→PP61280.2930.2350.266
CoBEVTSD2→PP81280.2930.2350.261
CoBEVTSD2→SD31280.2930.2350.268
CoBEVTSD3→PP61280.2930.2350.266
CoBEVTSD3→PP81280.2930.2350.261
CoBEVTSD3→SD21280.2930.2350.275
V2X-ViTPP6→PP81280.2930.2950.322
V2X-ViTPP6→SD21280.2930.2950.330
V2X-ViTPP6→SD31280.2930.2950.322
V2X-ViTPP8→PP61280.2930.2950.328
V2X-ViTPP8→SD21280.2930.2950.330
V2X-ViTPP8→SD31280.2930.2950.322
V2X-ViTSD2→PP61280.2930.2950.328
V2X-ViTSD2→PP81280.2930.2950.322
V2X-ViTSD2→SD31280.2930.2950.322
V2X-ViTSD3→PP61280.2930.2950.328
V2X-ViTSD3→PP81280.2930.2950.322
V2X-ViTSD3→SD21280.2930.2950.330
ParConPP6→PP81280.2930.3060.347
ParConPP6→SD21280.2930.3060.357
ParConPP6→SD31280.2930.3060.355
ParConPP8→PP61280.2930.3060.352
ParConPP8→SD21280.2930.3060.357
ParConPP8→SD31280.2930.3060.355
ParConSD2→PP61280.2930.3060.352
ParConSD2→PP81280.2930.3060.347
ParConSD2→SD31280.2930.3060.355
ParConSD3→PP61280.2930.3060.352
ParConSD3→PP81280.2930.3060.347
ParConSD3→SD21280.2930.3060.357
AttnFusePP6→PP6640.2930.2970.299
AttnFusePP8→PP8640.2930.2970.298
AttnFuseSD2→SD2640.2930.2970.307
AttnFuseSD3→SD3640.2930.2970.309
CoBEVTPP6→PP6640.2930.2350.238
CoBEVTPP8→PP8640.2930.2350.234
CoBEVTSD2→SD2640.2930.2350.257
CoBEVTSD3→SD3640.2930.2350.242
V2X-ViTPP6→PP6640.2930.2950.291
V2X-ViTPP8→PP8640.2930.2950.271
V2X-ViTSD2→SD2640.2930.2950.324
V2X-ViTSD3→SD3640.2930.2950.318
ParConPP6→PP6640.2930.3060.323
ParConPP8→PP8640.2930.3060.319
ParConSD2→SD2640.2930.3060.349
ParConSD3→SD3640.2930.3060.336
AuthorsOpenReview ↗

Response to Reviewer mbs5 (2/4)

W2/Q2: Weighted late fusion comparison [1/2]


Experimental setup

We added training-free weighted late fusion (WLF), using weighted box fusion, to the late-fusion (LF) comparison.

Standard LF/WLF uses all boxes. ALF-Box LF/WLF uses the Top-20 boxes selected by confidence score, following the same criterion as ALF.


Confidence threshold (Fig. 4)

Entries are [email protected]. Bold marks the highest value.

ConditionStandard LFALF-Box LFStandard WLFALF-Box WLFALF
0.10.1680.2580.1060.2350.321
0.20.2430.2580.2050.2360.322
0.30.2660.2580.2460.2370.322
0.40.2730.2580.2640.2400.321
0.50.2780.2560.2770.2470.320
0.60.2840.2510.2830.2490.315
0.70.2860.2440.2840.2470.306
0.80.2910.2440.2840.2460.299
0.90.2920.2430.2810.2440.298

ALF achieves the highest [email protected] at all nine confidence thresholds, outperforming both WLF variants. We will add the WLF comparisons for Fig. 4.


Detection-level corruptions (Fig. 5(a))

Entries are [email protected]. FP counts are injected boxes per AUX.

ConditionStandard LFALF-Box LFStandard WLFALF-Box WLFALF
Clean0.2430.2450.2070.2360.322
Box localization noise σ = 0.1 m0.2320.2360.2000.2270.317
Box localization noise σ = 0.3 m0.1830.1850.1540.1750.299
Box localization noise σ = 0.5 m0.1480.1520.1190.1370.279
Box heading noise σ = 0.2°0.2440.2450.2070.2360.319
Box heading noise σ = 0.6°0.2440.2450.2070.2360.319
Box heading noise σ = 1.0°0.2430.2450.2070.2360.320
Box dropout rate 25%0.2520.2450.2160.2370.314
Box dropout rate 50%0.2620.2440.2270.2370.308
Box dropout rate 75%0.2750.2430.2430.2380.302
Injected FP boxes: 50.1780.1570.1700.1590.259
Injected FP boxes: 100.1450.1170.1470.1210.224
Injected FP boxes: 200.1090.0680.1180.0680.169

ALF achieves higher [email protected] than both WLF variants in the clean condition and all twelve corrupted conditions. We will include these additional comparisons for Fig. 5(a) in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer mbs5 (3/4)

W2/Q2: System-level robustness [2/2]


System-level perturbations (Fig. 5(b))

Entries are [email protected]. Bold marks the highest value.

ConditionStandard LFALF-Box LFStandard WLFALF-Box WLFALF
Sensor localization noise σ = 0.1 m0.2310.2330.2010.2290.318
Sensor localization noise σ = 0.3 m0.1800.1830.1610.1800.300
Sensor localization noise σ = 0.5 m0.1460.1460.1350.1450.279
Sensor heading noise σ = 0.2°0.2300.2330.2000.2270.312
Sensor heading noise σ = 0.6°0.1950.1990.1760.1970.300
Sensor heading noise σ = 1.0°0.1740.1780.1590.1750.287
Latency 100 ms0.2180.2220.1880.2140.314
Latency 300 ms0.1940.1980.1680.1900.301
Latency 500 ms0.1820.1880.1570.1790.295

ALF achieves the highest [email protected] across all nine system-level conditions, outperforming both WLF variants. We will add the WLF comparisons for Fig. 5(b) in the revised manuscript.

Comments on the paper

1 comment

We thank the reviewers for their careful assessments and constructive feedback throughout the rebuttal. Their comments helped us strengthen the experimental comparisons, examine ALF’s limitations, and clarify its training and deployment requirements.


Strengths recognized by the reviewers

  • mbs5 recognized the well-motivated problem formulation, systematic evaluation across 64 zero-shot cases and four fusion backbones, communication-efficient design, and evidence supporting feature synthesis and robustness.
  • ziD9 highlighted the common box-message interface combined with feature-level interaction, adaptation-free integration of independently configured agents, and evaluation across multiple forms of heterogeneity.
  • wjK1 acknowledged the practical value of reducing new-agent adaptation costs, the compact 120-byte box payload, and the empirical evidence supporting feature synthesis.

Main concerns and improvements

ReviewerMain limitationImprovement
mbs5The practical benefit over no collaboration was unclear.Compared No Fusion, ALF (Ego-Only), and ALF across all 64 zero-shot cases.
mbs5, ziD9Stronger late-fusion baselines and direct evidence for feature synthesis were needed.Added LF comparisons with matched Top-20 AUX candidates and different message formats, together with weighted late-fusion robustness experiments.
ziD9, wjK1The channel-alignment procedure required clarification.Clarified channel alignment and projection bypass conditions for each baseline.
mbs5, ziD9EFS training and checkpoint reuse were insufficiently specified, and AUX1 results were inconsistent.Clarified training data, message generation, teacher construction, and checkpoint reuse, and corrected the AUX1 evaluation tables.
mbs5Camera-based collaboration had not been evaluated.Evaluated camera AUX zero-shot collaboration across four fusion backbones with No Fusion and feature-fusion baselines.
ziD9The efficiency–accuracy trade-off of quantization was unclear.Compared detection accuracy and computation time under EFS-only and broader model quantization, using QuantV2X’s computation-time notation.
wjK1Information loss from omitted box attributes was insufficiently examined.Compared six- and nine-field messages using class-wise AP and visibility-stratified recall.
wjK1Cross-dataset generalization and deployment costs needed further evidence.Evaluated V2XSet-trained AUX detectors on V2X-Real and measured inference time, GPU memory, and EFS runtime by collaborator count.
mbs5, wjK1The relationship to DiffAlign and GT-Space required clarification.Added methodological comparisons of feature spaces, transmitted information, and adaptation requirements.

Improvements through the rebuttal

  • Stronger comparisons — WLF, LF with different message formats, ALF (Ego-Only), and methodological comparisons with GT-Space and DiffAlign.
  • Efficiency and deployment analysis — model quantization, accuracy–computation trade-offs, and runtime and GPU-memory profiling.
  • Deeper box-message analysis — six- versus nine-field comparisons using class-wise AP and visibility-stratified recall.
  • Broader heterogeneity evaluation — camera auxiliaries and AUX detectors trained on a different dataset.

We have revised the main text and appendix to reflect the review comments and incorporated the additional experiments, corrections, and clarifications described above. We sincerely thank all reviewers, the AC, SACs, and PCs for their time and constructive contributions throughout the rebuttal process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.