Intermediate-fusion cooperative perception lets a vehicle detect objects beyond its own field of view by fusing features shared by nearby agents. When only a collaborator's cameras are degraded by fog, motion blur or exposure errors, the features it shares drift away from the training distribution while the ego input stays clean, and cooperation can become worse than not cooperating at all. Test-time batch-normalization adaptation, the standard training-free remedy, estimates statistics over the pooled batch of all agents and therefore normalizes both the clean ego and the corrupted collaborator with a mixture of two distributions. We show that this mixing harms even the clean ego: when the ego shares BN statistics with a corrupted collaborator, its detection from its own features drops by 3.58 to 6.68 [email protected] points although its cameras are untouched. We propose AgentNorm, an agent-wise test-time normalization that i) estimates the normalization statistics of every collaborator from its own current observation and ii) anchors the ego to its source statistics, so that the ego path remains exactly the original model. AgentNorm has no learnable parameters, needs no gradients or labels, is independent of the order of test frames, and runs locally on each agent without changing the transmitted features. On camera-based OPV2V with four fusion checkpoints and collaborator-only corruptions, AgentNorm raises [email protected] under severity-5 fog from 29.42-37.25 to 43.32-58.86 and outperforms pooled normalization by 5.53-12.94 points under fog and motion blur, at an extra latency of 2.3 ms per frame. On the real-world DAIR-V2X-C dataset, it outperforms pooled normalization in all ten infrastructure-corruption settings. We also report the cases it does not fix: occlusion, which removes image content rather than shifting feature statistics, and the case in which all agents share the same shift.
Figures & tables
Setting
Source data
Target labels
Re- training
Test-time gradients
Agent- specific shift
Fine-tuning
–
✓
✓
–
–
Domain adaptation
✓
–
✓
–
–
Domain generalization
✓
–
✓
–
–
Test-time training
✓
–
✓
✓
–
Fully TTA (e.g., Tent)
–
–
–
✓
–
Test-time BN
–
–
–
–
–
TABLE I: Adaptation settings. Our setting keeps the constraints of test-time BN and adds that the shift may affect only a subset of the agents processed by the same network.
Fig. 1: Statistics of the collaborator’s BEV features versus the severity of its camera corruption (AttFuse, OPV2V test, ego clean). (a) Mean normalized distance to the source statistics. (b) Mean local variance. Both change monotonically for all corruption types.
Corruption
Sev.
AUROC
No adapt.
Ego only
Score gating
Random gating
Oracle gating
Fog
3
0.719
38.85
38.31
39.09
39.50
41.16
Fog
5
0.743
37.25
38.31
38.45
38.80
40.02
Motion blur
3
0.708
42.53
38.31
40.36
42.06
44.16
Motion blur
5
0.752
36.48
38.31
38.14
38.25
39.91
Brightness
3
0.650
53.00
38.31
44.35
50.78
51.78
Brightness
5
0.719
42.37
38.31
41.01
41.69
42.28
TABLE II: Region-level gating of the collaborator (AttFuse, OPV2V, [email protected]). Only the collaborators’ cameras are corrupted. AUROC: localization of shifted cells by the gating score. Gating columns report the best of k∈{10,30,50}% . Underlined: cooperation below ego-only. Oracle uses the paired clean frame and is not a feasible method.
Fig. 2: One BN layer of the per-agent encoder with a clean ego ( X0 , blue) and a corrupted collaborator ( X1 , orange). (a) Pooled test-time BN normalizes both agents with the mixture statistics (μP,σP2) , which shifts the clean ego (dashed) and under-corrects the collaborator. (b) AgentNorm keeps the ego on the source statistics (ego anchoring) and normalizes the collaborator with the statistics of its own current observation (agent-wise statistics).
Fig. 3: Mechanism analysis on AttFuse (OPV2V, severity 5). (a) [email protected] of the clean ego detected from its own feature alone, under source statistics, its own statistics, and statistics pooled with a clean or corrupted collaborator. Pooling with a corrupted collaborator lowers the ego’s own detection by 3.58–6.68 points. (b) [email protected] of cooperative detection when agent-wise statistics are applied only in the BEV backbone, only in the image encoder, or in both (AgentNorm).
Fig. 4: BEV features on an OPV2V test frame (AttFuse) with the collaborator under severity-5 fog and a clean ego. (a)–(c) Channel-averaged magnitude of the collaborator’s feature with clean cameras, with fog, and with fog after AgentNorm. (d) Channel-averaged absolute change of the ego’s feature relative to the source model under Pooled BN. Under AgentNorm the same change averages 7.3×10−5 (floating-point level, against 0.016 for Pooled BN) because the ego keeps its source statistics, so it is not plotted.
Fig. 5: One collaborator camera image of OPV2V under the four corruptions at severity 5. Only the collaborators’ cameras are corrupted; the ego’s cameras stay clean.
Fig. 6: [email protected] of AttFuse on OPV2V as the severity of the collaborator’s corruption increases (ego clean). The dotted line is the ego-only model. Under fog and motion blur, AgentNorm degrades much more slowly than Pooled BN and stays above the ego-only model; under occlusion, no method helps.
Severity 3
Severity 5
Corruption
None
Pooled
Ours
None
Pooled
Ours
Fog
1.65
2.78
6.75
1.50
2.07
5.88
Snow
1.57
4.82
7.35
1.16
2.44
5.17
Frost
1.41
3.42
6.00
1.30
2.71
5.40
Brightness
5.61
7.25
8.82
1.17
3.79
6.08
Motion blur
8.97
7.55
8.93
6.86
6.01
7.46
TABLE IV: Roadside-only corruption on DAIR-V2X-C (camera CoAlign, [email protected] in %). For each severity: no adaptation / Pooled BN / AgentNorm. Best per severity in bold. The vehicle-only model reaches 1.65.
Fig. 7: Remedies for a corrupted collaborator (AttFuse, OPV2V, collaborator fog at severity 5). The dashed line marks no adaptation. The oracle uses ground-truth labels and is shown only as a reference.
Fig. 8: Detections on an OPV2V test frame (AttFuse) with two collaborators (orange triangles) under severity-5 fog and a clean ego (black triangle). From left to right: no adaptation, Pooled BN and AgentNorm. Solid green boxes: ground truth; dashed boxes: predictions. Gray points are the LiDAR scan provided by OPV2V, drawn only to show the road layout and the agents’ surroundings; no model in this paper uses LiDAR. True positives at IoU 0.5: 6, 8 and 13 of 21 objects. The frame was selected as the one with the largest gain of AgentNorm over no adaptation among 115 candidate frames (every 18th test frame); statistics over all candidates are given in the text.
Variant
Fog
Motion blur
Brightness
Occlusion
Clean
No adaptation
37.25 / 19.04
36.48 / 18.59
42.37 / 21.46
39.92 / 19.87
56.81 / 27.03
Pooled BN
40.65 / 18.08
42.90 / 18.94
53.87 / 23.47
35.17 / 14.83
59.17 / 27.25
Agent-wise, symmetric
53.11 / 23.55
50.69 / 22.24
56.60 / 25.20
37.06 / 17.48
58.76 / 26.69
Agent-wise, ego-anchored (AgentNorm)
51.98 / 24.35
49.83 / 22.88
55.59 / 25.70
35.63 / 17.95
57.85 / 27.18
AgentNorm, EMA statistics ( m=0.05 )
51.12 / 24.17
–
–
–
–
Oracle: GT-supervised collab. BN †
57.13 / 26.75
–
–
–
–
TABLE V: Ablation on AttFuse (OPV2V). Only the collaborators are corrupted (severity 5) or nothing is corrupted (clean). Entries are [email protected] / [email protected] in %. † Uses ground-truth labels; reference only.
Corruption
No adapt.
Pooled BN
Agent-wise (sym.)
Fog
2.44 / 1.63
41.74 / 15.39
41.71 / 15.54
Motion blur
0.66 / 0.33
32.50 / 10.31
33.34 / 10.11
Brightness
11.45 / 5.26
47.98 / 20.12
46.97 / 19.16
Occlusion
2.31 / 0.83
6.25 / 3.00
6.07 / 3.11
TABLE VI: Ego and collaborators corrupted with the same corruption (AttFuse, OPV2V, severity 5). [email protected] / [email protected] in %.
Method
Mean
Median
P95
Memory
No adaptation
254.32
184.38
495.70
2669
No adaptation (repeat)
255.19
185.33
496.40
2669
Pooled BN
255.28
184.73
497.33
2669
AgentNorm (ours)
256.65
186.16
501.99
2669
TABLE VII: Per-frame latency (ms) and peak GPU memory (MB) of AttFuse on OPV2V, collaborator fog at severity 5, including post-processing. 200 frames, the first 20 discarded, one RTX 3090.
Cooperative perception lets agents share information to expand coverage and improve scene understanding. However, in real-world scenarios, diverse and unpredictable corruptions undermine its robustness and generalization. To address these challenges, we introduce CoopDiff, a diffusion-based cooperative perception framework that mitigates corruptions via a denoising mechanism. CoopDiff adopts a teacher-student paradigm: the Quality-Aware Teacher performs voxel-level early fusion with Quality of Interest weighting and semantic guidance, then produces clean supervision features via a diffusion denoiser. The Dual-Branch Diffusion Student first separates ego and cooperative streams in encoding to reconstruct the teacher's clean targets. And then, an Ego-Guided Cross-Attention mechanism facilitates balanced decoding under degradation by adaptively integrating ego and cooperative features. We evaluate CoopDiff on two constructed multi-degradation benchmarks, OPV2Vn and DAIR-V2Xn, each incorporating six corruption types, including environmental and sensor-level distortions. Benefiting from the inherent denoising properties of diffusion, CoopDiff consistently outperforms prior methods across all degradation types and lowers the relative corruption error. Furthermore, it offers a tunable balance between precision and inference efficiency.
Gong Chen, Chaokun Zhang, Pengcheng Lv
School of Computer Science and Technology, Tianjin University · School of Cybersecurity, Tianjin University · School of Future Technology, Tianjin University
Multi-agent collaborative perception enables autonomous systems to overcome individual sensing limits through collective intelligence. However, real-world sensor and communication corruptions severely undermine this advantage. Crucially, existing approaches treat corruptions as static perturbations or passively conform to corrupted inputs, failing to actively recover the underlying clean semantics. To address this limitation, we introduce Diff-KD, a framework that integrates diffusion-based generative refinement into teacher-student knowledge distillation for robust collaborative perception. Diff-KD features two core components: (i) Progressive Knowledge Distillation (PKD), which treats local feature restoration as a conditional diffusion process to recover global semantics from corrupted observations; and (ii) Adaptive Gated Fusion (AGF), which dynamically weights neighbors based on ego reliability during fusion. Evaluated on OPV2V and DAIR-V2X under seven corruption types, Diff-KD achieves state-of-the-art performance in both detection accuracy and calibration robustness.
Pengcheng Lyu, Chaokun Zhang, Gong Chen +2
School of Future Technology, Tianjin University, Tianjin, China · School of Cybersecurity, Tianjin University, Tianjin, China · School of Computer Science and Technology, Tianjin University, Tianjin, China
Most existing heterogeneous cooperative perception methods depend on prior preparation like offline joint training or tailored collaborator-model adaptation. Such preprocessing is, however, generally impractical in real scenarios, as agents are usually independently trained by different developers and meet occasionally online. This work investigates preparation-free heterogeneous cooperative perception, where agents use independently trained single-agent detectors without any pre-deployment coordination. We find direct cross-agent fusion under this setting greatly underperforms ego-only perception. We present BOLT, a lightweight plug-and-play module that adapts neighboring features online via ego-as-teacher distillation, requiring only ego predictions without ground-truth labels. BOLT leverages high-confidence ego perception features to guide cross-agent feature-domain alignment, while enabling neighbors to contribute features in the ego's low-confidence regions. With only 0.9M trainable parameters, BOLT improves AP@50 by up to 32.3 points over vanilla unadapted fusion in the preparation-free setting. It consistently outperforms ego-only results on DAIR-V2X and OPV2V, across different encoder pairs and fusion strategies. Code: https://github.com/sidiangongyuan/BOLT.
Kang Yang, Tianci Bu, Peng Wang +2
Renmin University of China · National University of Defense Technology