Changes in sensor height and viewpoint alter object-level point distributions, making cross-platform LiDAR unsupervised domain adaptation (UDA) difficult. Self-training uses labeled source scans and unlabeled target scans, yet a retained prediction may provide a useful target location while enclosing sparse foreground returns, background clutter, or points inconsistent with the predicted box. We refer to this mismatch as box-point inconsistency. We introduce SimFuse3D, which preserves the target placement and repairs the associated pseudo object using measured geometry from labeled source scans. Object Memory retrieves a similar labeled source instance. Target Simulation places the retrieved source geometry at the target location, aligns its points with the target viewing geometry, and filters the aligned crop to approximate the target observation. Confidence-Guided Multi-Stage Localization Reweighting (CMLR) maps each target pseudo-object confidence score to a bounded weight shared by RPN localization and R-CNN box regression. All components operate only during adaptation, leaving the detector architecture and inference graph unchanged. Across six cross-platform transfers, SimFuse3D consistently outperforms Pi3DET-Net and achieves the best performance among the compared adaptation methods on nearly all metrics. On nuScenes-to-KITTI, it ranks first among the compared adaptation methods with both evaluated detectors.
Figures & tables
Fig. 1: Motivation and qualitative comparison. (a) Pi3DET-Net [ 3 ] retains the target point observations enclosed by accepted pseudo boxes, allowing unreliable geometry to enter localization supervision. (b) Our method, SimFuse3D, retrieves measured source geometry, aligns it with the target placement and viewpoint, and reweights target localization supervision with CMLR. (c) Vehicle → Drone example. The upper dashed green circle marks one ground-truth vehicle, and the lower circle marks two adjacent vehicles. All three boxes contain only a few LiDAR returns. Pi3DET-Net misses the three vehicles in the two marked regions, whereas SimFuse3D detects them. Red boxes show ground truth, blue boxes show predictions, and purple points are returns inside the ground-truth boxes.
Fig. 2: Overview of our SimFuse3D framework . (a) The detector is initialized by source-domain pre-training, and (b) target inference generates pseudo labels for adaptation. During adaptation, (c) Object Memory retrieves labeled source instances for target pseudo objects, (d) Target Simulation repairs target pseudo objects with retrieved source geometry, and (e) CMLR reweights localization supervision according to pseudo-object confidence. The modules in (c)–(e) are used only during adaptation and are absent from the deployed inference graph.
Fig. 3: Object Memory and the proposed Target Simulation . Object Memory stores source points, ground-truth boxes, metadata, and fixed Point-NN [ 21 ] descriptors for source-instance retrieval. A target pseudo object queries the memory for a labeled source instance. Target Simulation uses the target pseudo object to guide viewpoint alignment and point-distribution simulation, producing the simulated target observation.
Source Domain
Method
Pi3DET (Quadruped)
Pi3DET (Drone)
PV-RCNN
Voxel R-CNN
PV-RCNN
Voxel R-CNN
AP@0.7
AP@0.5
AP@0.7
AP@0.5
AP@0.7
AP@0.5
AP@0.7
AP@0.5
nuScenes
Source Only
40.11 / 31.23
43.95 / 42.00
41.84 / 33.39
45.27 / 43.36
35.56 / 24.96
38.36 / 35.76
37.84 / 24.21
44.53 / 39.60
ST3D [ 4 ]
54.83 / 42.66
58.94 / 56.80
50.91 / 41.38
53.25 / 51.69
51.56 / 31.54
58.35 / 51.54
52.20 / 33.53
56.85 / 51.29
ST3D ‡ [ 4 ]
54.68 / 43.34
58.64 / 57.78
50.46 / 40.95
53.02 / 51.20
50.94 / 31.84
57.68 / 51.07
51.95 / 33.25
56.65 / 51.14
ST3D++ [ 5 ]
56.08 / 41.90
58.84 / 58.03
51.05 / 41.45
53.76 / 51.70
51.26 / 33.35
56.63 / 51.34
54.32 / 34.70
59.41 / 54.03
TABLE I: Cross-platform adaptation to Pi3DET Quadruped and Drone. Entries are APBEV/AP3D (%). Entries within each detector column share the detector and evaluation configuration; Pi3DET-Net is reproduced under this setup. ST3D/ST3D++ use random object scaling unless ‡ marks no ROS. “Target Platform” is fully supervised; the best scores among the listed adaptation methods are bolded.
Task
Method
PV-RCNN
Voxel R-CNN
AP@0.7
AP@0.5
AP@0.7
AP@0.5
Q → D
Source Only
27.52 / 11.76
32.60 / 27.62
28.02 / 12.14
33.93 / 28.54
ST3D [ 4 ]
31.14 / 13.93
38.27 / 31.90
30.31 / 20.06
36.03 / 33.05
ST3D++ [ 5 ]
34.35 / 16.46
41.97 / 36.98
33.88 / 20.89
39.29 / 36.31
ReDB [ 12 ]
27.71 / 16.79
33.57 / 29.66
32.21 / 22.57
36.80 / 34.13
MS3D++ [ 6 ]
31.77 / 16.47
40.00 / 35.27
26.68 / 19.43
29.41 / 29.20
TABLE II: Bidirectional adaptation between Pi3DET Quadruped (Q) and Drone (D). Entries are APBEV/AP3D (%). Entries within each detector column share the detector and evaluation configuration; Pi3DET-Net is reproduced under this setup. “Target Platform” is fully supervised; the best scores among the listed adaptation methods are bolded.
Method
APBEV
AP3D
Source Only
64.98
38.85
SN [ 7 ]
55.00
43.95
ST3D [ 4 ]
74.47
50.24
ST3D [ 4 ] w/ SN
78.40
70.90
ST3D++ [ 5 ]
81.70
45.35
ST3D++ [ 5 ] w/ SN
84.98
75.50
TABLE III: nuScenes → KITTI adaptation with PV-RCNN at IoU 0.7 . Entries are AP (%); the best adaptation scores are bolded. “w/ SN” denotes statistical normalization, and “Target Platform” is fully supervised.
Method
APBEV
AP3D
Source Only
41.58
17.62
SN [ 7 ]
33.95
23.26
ST3D [ 4 ]
77.29
35.58
ST3D [ 4 ] w/ SN
87.11
66.02
ST3D++ [ 5 ]
82.08
38.10
ST3D++ [ 5 ] w/ SN
85.84
68.64
TABLE IV: nuScenes → KITTI adaptation with Voxel R-CNN at IoU 0.7 . Entries are AP (%); the best adaptation scores are bolded. Unreported methods are omitted, and “Target Platform” is fully supervised.
TS
CMLR
Pi3DET (Quadruped)
Pi3DET (Drone)
AP@0.7
AP@0.5
AP@0.7
AP@0.5
58.76 / 47.44
63.68 / 61.22
65.82 / 49.59
71.22 / 65.57
✓
60.94 / 48.22
65.97 / 63.58
67.53 / 49.66
70.94 / 67.36
✓
✓
62.54 / 49.34
67.71 / 65.40
67.58 / 50.77
73.05 / 69.35
TABLE V: Component ablation with Voxel R-CNN on the Vehicle-source transfers. TS and CMLR denote Target Simulation and Confidence-Guided Multi-Stage Localization Reweighting. Entries are APBEV/AP3D (%); best scores are bolded.
Task
GT pts.
#GT
Dist. (m)
R@0.5
R@0.7
V → Q
0–15
304
28.24
9.54/ 11.51 (+1.97)
4.93/ 5.26 (+0.33)
16–30
180
25.16
45.56/ 51.11 (+5.56)
27.22/ 30.00 (+2.78)
≥31
1,581
18.46
80.83/ 86.08 (+5.25)
66.86/ 70.46 (+3.61)
All
2,065
20.41
67.26/ 72.06 (+4.79)
54.29/ 57.34 (+3.05)
V → D
0–15
1,573
36.10
29.37/ 36.62 (+7.25)
16.59/ 19.64 (+3.05)
16–30
871
28.77
63.15/ 65.21 (+2.07)
42.37/ 44.32 (+1.95)
TABLE VI: Voxel R-CNN recall by target ground-truth point support. Entries are Pi3DET-Net/Ours ( Δ ), in percent; V, Q, and D denote Vehicle, Quadruped, and Drone. #GT is the object count, Dist. is the median sensor distance, and Δ is computed before rounding.
Fig. 4: Pseudo-label reliability on Vehicle → Drone. Each cell reports Q@IoU≥0.5 (%) and the sample count n for a teacher-confidence and in-box point-support bin.
Fig. 5: Vehicle → Drone comparison. Orange outlines mark three sparse vehicles recovered by SimFuse3D at IoU 0.5 but missed by Pi3DET-Net.
Reliable 3D perception of vulnerable road users (VRUs) such as cyclists and pedestrians is essential for their safety in urban traffic and a core requirement for autonomous driving (AD). Alongside advances in vehicle-based perception, research increasingly equips bicycles with sensors to study traffic from a perspective native to VRUs. Such platforms still rely on LiDAR detectors originally trained on vehicle data, yet annotated 3D data from a cyclist's perspective is scarce. How well these detectors generalise to this setting has not been evaluated. We present a 3D object detection benchmark of 1,027 annotated LiDAR keyframes (over 18,000 3D bounding boxes) from the FUSE-Bike platform in urban Munich. We evaluate four nuScenes-pre-trained detectors against 1,854 human-verified ground-truth (GT) boxes both in their original form and after finetuning on training labels produced by a VRU-dedicated auto-labelling pipeline that requires no manual annotation. The zero-shot domain gap is concentrated on the VRU classes. Finetuning recovers most of it, improving mean average precision (mAP) by up to 23.4 points with the largest gains on pedestrians and cyclists, and the adapted detectors even surpass the quality of the auto-labels they were trained on. The benchmark provides a reproducible baseline for VRU-centric 3D detection and shows that auto-labels are a viable substitute for manual annotation when adapting vehicle-trained detectors to a cyclist platform.
Mario Finkbeiner, Max A. Buettner, Kanak Mazumder +1
Intelligent Vehicles Lab (IVL), Munich University of Applied Sciences, Munich, Germany
We present our solution to the LUMPI track of the UCF UrbanTwin Sim2Real LiDAR Challenge at the 6th DriveX Workshop, ECCV 2026. The detector must be trained only on synthetic data and is evaluated on 50 held-out real LiDAR frames; a separate 50-frame synthetic submission is evaluated for point-cloud realism. Our method addresses the Sim2Real gap at three levels. First, we align synthetic scans to the 50k-point test density and build a 30k-record training pool using UT-LUMPI geometry, RangeLDM-based sampling diversification, rare-class copy-paste, and pedestrian-oriented augmentation. Second, complementary DSVT detectors and Car/Bus PointPillars specialists are trained under the same synthetic-only constraint. Third, predictions are integrated by class-aware routing, asymmetric agreement fusion, constrained residual-recall supplementation, class-coverage auditing, and selective box-size calibration. The realism branch is optimized independently with radial-density matching, weak affine calibration, and calibrated set mixing. The final submission obtains a Combined Score of 0.4692, a Detection Score of 0.1797, a Realism Score of 0.9035, and 3D mAP@0.5 of 0.1258.
With the advancement of autonomous driving, numerous annotated multi-modality datasets have become available. This presents an opportunity to develop domain-adaptive 3D object detectors for new environments without relying on labor-intensive manual annotations. However, traditional domain adaptation methods typically focus on a single source domain or a single modality, limiting their effectiveness in multi-source, multi-modality scenarios. In this paper, we propose a novel framework for multi-source, multi-modality unsupervised domain adaptation in 3D object detection for autonomous driving. Given multiple labeled source domains and one unlabeled target domain, our framework first introduces hierarchical spatially-conditioned (HSC) domain classifiers, which jointly align features from both camera and LiDAR modalities at two distinct levels for each source-target domain pair. To effectively leverage information from multiple source domains, we construct a prototype graph between each pair of domains. Based on this, we develop a prototype graph weighted (PGW) multi-source fusion strategy to aggregate predictions from multiple source detection heads. Experimental results on three widely used 3D object detection datasets - Waymo, nuScenes, and Lyft - demonstrate that our proposed framework effectively integrates information across both modalities and source domains, consistently outperforming state-of-the-art methods.
Xiaohu Lu, Hamed Khatounabadi, Hayder Radha
Electrical and Computer Engineering Michigan State University East Lansing, USA