LiDAR has become an essential sensing modality in autonomous driving, robotics, and smart-city applications. However, ghost points (or ghosts), which are false reflections caused by multi-path laser returns from glass and reflective surfaces, severely degrade 3D mapping and localization accuracy. Prior ghost removal relies on geometric consistency in dense point clouds, failing on mobile LiDAR's sparse, dynamic data. We address this by exploiting full-waveform LiDAR (FWL), which captures complete temporal intensity profiles rather than just peak distances, providing crucial cues for distinguishing ghosts from genuine reflections in mobile scenarios. As this is a new task, we present Ghost-FWL, the first and largest annotated mobile FWL dataset for ghost detection and removal. Ghost-FWL comprises 24K frames across 10 diverse scenes with 7.5 billion peak-level annotations, which is 100x larger than existing annotated FWL datasets. Benefiting from this large-scale dataset, we establish a FWL-based baseline model for ghost detection and propose FWL-MAE, a masked autoencoder for efficient self-supervised representation learning on FWL data. Experiments show that our baseline outperforms existing methods in ghost removal accuracy, and our ghost removal further enhances downstream tasks such as LiDAR-based SLAM (66% trajectory error reduction) and 3D object detection (50x false positive reduction). The dataset and code is publicly available and can be accessed via the project page: https://keio-csg.github.io/Ghost-FWL
Figures & tables
Access & Platform
Sensor
Labels
Dataset
Year
Public
Platform
FWL
LiDAR Dim.
Ray Den.
Ghost
FWL Data
Frames/ Scenes †
Annotated Peaks
UNIST [ 52 ]
2017
✓
Stationary
✗
3D
278
✓
✗
–
–
Leddar PixSet [ 9 ]
2021
✓
Mobile
✓
3D
0.267
✗
✓
–
–
Lee et al. [ 23 ]
2023
✗
Stationary
✗
3D
278
✗
✗
–
–
FRACTAL [ 13 ]
2024
✓
Aerial
✗
2D
–
✗
✗
–
–
Scheuble et al. [ 38 ]
2025
✗
Mobile
✓
3D
2.56
✗
✓
0.24k / 2
NA
Table 1: Comparison of LiDAR real-world datasets for ghost detection and/or full-waveform analysis. Our Ghost-FWL contains mobile LiDAR full-waveform measurements and is one hundred times larger than prior work, making it the largest annotated FWL dataset.
Figure 2: Overview of the Ghost-FWL. Left: Our dataset includes both indoor and outdoor scenes. Based on the dense 3D maps as shown in Scene, we annotated FWL data with semantic labels: Ghost (red), Object (green), Glass (blue), Noise . Gray regions are excluded from annotation. Right shows the data acquisition setup and dataset statistics, including the incident angle distribution and LiDAR positions examples. Data were collected at three different times of day: Morning (AM10–12), Daytime (PM12–5), and Evening (PM5–7).
Figure 3: FWL-based ghost removal framework. Given FWL data, our framework predicts and removes ghost-related signals. Our model consists of a transformer-based encoder and an MLP head. We further introduce FWL-MAE, a masked autoencoder designed for representation learning on FWL data, explicitly trained to reconstruct peak position , amplitude and width . The ghosts detected by our model are then removed from FWL data, and the cleaned data are utilized for downstream tasks such as SLAM and 3D object detection.
Figure 4: Peak classification results and point cloud visualization after applying ghost removal. All results were obtained using the proposed framework. Red, green and blue indicates Ghost , Object , Glass , respectively.
Method
Recall ( ↑ )
Removal Rate ( ↑ )
MARMOT [ 39 ]
0.746
0.910
Ours w/o FWL-MAE
0.704
0.900
Ours
0.751
0.918
Table 2: Comparison of ghost removal performance with other methods.
Method
ATE [m] ( ↓ )
RTE [m] ( ↓ )
Dual-Peak [ 32 , 17 ]
0.715 ± 0.433
0.741 ± 0.406
Multi-Peak [ 27 ]
1.547 ± 1.394
1.602 ± 1.381
Ours
0.245 ± 0.138
0.245 ± 0.131
Table 3: SLAM performance under different LiDAR signal preprocessing methods.
Figure 5: Trajectory and mapping generated by SLAM using Multi-Peak processing (left) and our ghost removal method (right). Multi-Peak processing includes numerous ghost points in the reconstructed map, leading to trajectory drift. The proposed method yields a trajectory that more closely follows the ground-truth path (white) by effectively removing ghost artifacts.
Method
Ghost FP Rate [%] ( ↓ )
Dual-Peak [ 32 , 17 ]
75.8
Multi-Peak [ 27 ]
67.9
Ours
1.34
Table 4: Ghost-induced false positives in ghost detection.
Figure 6: Qualitative evaluation of 3D object detection with Multi-Peak processing (left) and our ghost removal (right). Green bounding boxes indicate persons. With Multi-Peak, a ghost person is detected behind the glass wall, whereas our method suppresses this false detection.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Scene
Frames
Location
001
2500
Indoor
002
2500
Indoor
003
2749
Indoor
004
1853
Indoor
005
2500
Outdoor
006
2445
Outdoor
Appendix
Table 5: Ghost-FWL statistics by scene.
Figure 7: All scenes in Ghost-FWL. Our dataset includes both indoor and outdoor scenes. Based on the dense 3D maps as shown in Scene, we annotated FWL data with semantic labels: Ghost (red), Object (green), Glass (blue), Noise . Gray regions are excluded from annotation. Red crosses indicate the LiDAR positions during data acquisition. The RGB images show the scenery of the capture locations.
Scene
Threshold τ
# Glass Areas
001
0.5
8
002
0.5
48
003
0.5
10
004
0.5
9
005
0.5
15
006
0.5
5
Appendix
Table 6: Annotation parameters for each scene in the Ghost-FWL dataset. The threshold τ is the nearest-neighbor distance used to distinguish Object and Ghost points, and “# Glass Areas” denotes the number of manually annotated glass regions G .
Figure 8: Example of LiDAR position and perspective.
Figure 9: Example of 1 frame FWL data(top) and 50 × accumulation FWL data (bottom).
Figure 10: Example of raw annotation (top) and processed annotation for training (bottom). The circles indicate the annotation points for each class: Object (green), Glass (blue) and Ghost (red).
Name
Layer setting
Output dimension
Patch Embedding
3D Conv
B×Npatch×Dencoder
Flatten + Transpose
Positional Encoding
Sinusoidal Position Encoding (PE)
B×Npatch×Dencoder
Masking
Select Unmasked Tokens + PE
B×Nunmasked×Dencoder
Encoder
Transformer Blocks × 6
(pretraining) B×Nunmasked×Dencoder
LayerNorm
(fine-tuning) B×Npatch×Dencoder
Appendix
Table 7: FWL-MAE Architecture.
Method
Recall ( ↑ )
3D U-Net [ 8 ]
0.391
Lindell et al. [ 25 ]
0.641
MARMOT [ 39 ]
0.746
Ours w/o FWL-MAE
0.704
Ours
0.751
Appendix
Table 8: Comparison of ghost removal performance with other methods.
Figure 11: Peak classification results. Red, green and blue indicates Ghost , Object and Glass , respectively.
Train Data
Method
Recall ( ↑ )
100%
Ours w/o FWL-MAE
0.704
Ours
0.751 (+ 0.047)
70%
Ours w/o FWL-MAE
0.602
Ours
0.692 (+ 0.090)
50%
Ours w/o FWL-MAE
0.403
Ours
0.603 (+ 0.200)
Appendix
Table 9: Ablation study of FWL-MAE.
Method
Params [M]
FLOPs [G]
FPS †
FPS ‡
3D U-Net [ 8 ]
90.3
7610
6.6
0.45
Lindell et al. [ 25 ]
1.8
4090
7.0
0.48
Ours
194.1
23.5
32.1
2.35
Appendix
Table 10: Computational cost and inference time. FPS † denotes the inference speed for model input size (H,W,T)=(128,128,256) . FPS ‡ denotes the inference for full-frame size (H,W,T)=(332,400,256) .
Figure 12: Our model’s ghost detection on non-glass surfaces
Raw
Statistical Outlier Filter
Radius-based Outlier Filter
Dual-Peak ATE [m] ( ↓ )
1.248 ± 0.947
0.639 ± 0.573
0.503 ± 0.593
Multi-Peak ATE [m] ( ↓ )
2.489 ± 1.432
1.232 ± 0.928
0.887 ± 1.079
Ours ATE [m] ( ↓ )
0.294 ± 0.244
0.248 ± 0.171
0.328 ± 0.303
Dual-Peak RTE [m] ( ↓ )
1.221 ± 0.928
0.608 ± 0.529
0.471 ± 0.555
Multi-Peak RTE [m] ( ↓ )
2.513 ± 1.440
1.280 ± 0.901
0.832 ± 1.015
Ours RTE [m] ( ↓ )
0.288 ± 0.237
0.243 ± 0.157
0.319 ± 0.291
Appendix
Table 11: SLAM performance ablation with different point-cloud processing methods.
Figure 13: Trajectory and mapping results using Dual-Peak processing (left), Multi-Peak processing (center), and our ghost-removal method (right). The bottom image shows the scenery of the SLAM evaluation, recorded in a corridor enclosed by glass railings and doors. The scene is same to that used in the main paper Fig 5 , with the RGB scene image and the visualization results for the Dual-Peak method additionally included.
Scene
Indoor/ Outdoor
Pedestrians
Frames
PedScene 001
Indoor
2
14
PedScene 002
Outdoor
3
31
PedScene 003
Indoor
3
32
PedScene 004
Indoor
1
25
TOTAL
Indoor 3 / Outdoor 1
1 / 2 / 3
102
Appendix
Table 12: Details of the object-detection test dataset (PedScene).
Figure 14: Example scenes from the PedScene dataset used for 3D object detection. The dataset includes both indoor and outdoor environments with glass surfaces, building entrances, and glass-walled sidewalks.
Point cloud anomaly detection identifies and localizes unusual 3D geometry or surface defects, often in industrial inspection with few labeled failures.
3D object detection from LiDAR point clouds is a core problem in autonomous driving. Recent advances in self-supervised learning (SSL) enable scalable pretraining and transfers well to per-point tasks such as semantic and panoptic segmentation, but transfer to 3D detection remains weaker. We analyze recent SSL methods and find that most objectives are defined only on measured LiDAR returns from visible surfaces, leaving occluded and unobserved regions unconstrained. This visible-surface bias can be sufficient for point-wise prediction, but 3D detection requires robustness to missing structure. To address this gap, we propose GhostPoint, an SSL framework that hallucinates latent features in local neighborhoods around discovered instances, generated via a novel instance voxel dilation. In GhostPoint, an encoder processes observed returns, and an additional predictor infers neighborhood representations from observed context. In addition to standard encoder-level supervision, we introduce a predictor-level supervision scheme on sampled voxels from generated neighborhoods. Specifically, observed (visible/masked) voxels match teacher-encoder targets, while unobserved voxels match teacher-predictor hallucinations. This design encourages the learned representation to explicitly model structure beyond observed returns. Extensive evaluations on nuScenes and Waymo demonstrate that our method achieves state-of-the-art performance, consistently improving downstream 3D detection, especially under sparse scans and limited labels.
Mohamed Abdelsamad, Bin Yang, Michael Ulrich +4
Bosch Center for AI · University of Freiburg · University of Luebeck
Modern LiDARs are rapidly transitioning from bulky, mechanically scanned systems to ultra-compact, low-cost, solid-state arrays. This miniaturization-while enabling scalability, affordability, and camera-like data structures-introduces a new and severe failure mode: internal-multipath glare. When light from a bright or retroreflective surface reflects and scatters within the LiDAR, light that should reach a single pixel spreads across the pixel array. The resulting artifacts create phantom objects, obscure real ones, and produce safety-critical "ghosts in the point clouds." This paper introduces a physically grounded sensing model and algorithmic techniques for addressing this effect. We show that internal glare can be represented as a linear, scene-independent operator-the Transient Glare Spread Function (TGSF)-acting on the transient measurements. Building on this model, we develop a training-free approach that operates on low-level LiDAR detections (or echoes) prior to point-cloud formation, leveraging knowledge of the glare spread function to reason about the likelihood of each detection arising from glare. The resulting approach is compatible with existing LiDAR signal-processing pipelines, and deployable on unmodified commercial sensors. Using experiments with real single-photon LiDAR hardware, we demonstrate substantial suppression of severe glare artifacts while preserving true scene structure.
FMCW LiDAR measures per-point radial Doppler velocity alongside range, providing a motion cue unavailable in conventional time-of-flight sensors. Exploiting this signal at long range remains understudied. We present an FMCW LiDAR dataset of 575 sequences (57.5K frames) with over 8 million annotated 3D boxes across 16 detection classes and per-point labels across 24 semantic classes, captured by six commercial FMCW LiDAR sensors and six paired 4K cameras across eight Bay Area cities, including 237 nighttime sequences, with annotations extending to 400m. We define a benchmark with three tasks: 3D object detection, scene flow estimation, and semantic segmentation. Detection and scene flow are evaluated across three range bins to 400m, with a public evaluation server. We explore the impact of Doppler measurements on flagship recognition tasks, and find significant improvements up to 2X in detection AP of far-away vehicles and pedestrians, particularly in low-latency single-frame settings. We similarly find scene flow accuracy is significantly improved with Doppler measurements across all ranges. Our dataset and benchmark have been publicly released at https://scenes.aeva.com.