ALICE-LRI: A General Method for Lossless Range Image Generation for Spinning LiDAR Sensors without Calibration Metadata
Authors: Samuel Soutullo, Miguel Yermo, David L. Vilariño, Óscar G. Lorenzo, José C. Cabaleiro, Francisco F. Rivera
Organizations: Centro Singular de Investigación en Tecnoloxías Intelixentes (CiTIUS), Rúa de Jenaro de la Fuente Domínguez, Santiago de Compostela, 15782, A Coruña, Spain · Departamento de Electrónica e Computación, Universidade de Santiago de Compostela, Rúa Lope Gómez de Marzoa, Santiago de Compostela, 15782, A Coruña, Spain
3D LiDAR sensors are widely used in autonomous navigation, environmental monitoring, and precision mapping. Their point clouds are often projected into 2D range images for efficient processing, but conventional projection methods can lose points because of geometric inconsistencies. We present ALICE-LRI (Automatic LiDAR Intrinsic Calibration Estimation for Lossless Range Images), a metadata-free method for lossless range-image generation from calibrated spinning multi-beam LiDAR point clouds. It requires neither factory calibration files, packet-level data, nor beam identifiers, and targets operating regimes where scanlines contain enough returns for reliable parameter estimation. The method estimates the intrinsic calibration parameters relevant to projection, including beam configuration, angular distributions, and per-beam geometric corrections. Evaluation on the complete KITTI and DurLAR datasets yielded zero lost points across all point clouds. Reconstruction deviations remain below manufacturer-reported range-performance specifications, while the method runs in real time. A compression case study also shows improved reconstruction quality and a better quality-compression trade-off for a downstream application. These results enable range-image-based LiDAR processing in applications that require high measurement fidelity and preservation of acquired geometric information.
Figures & tables
Figure 1: LiDAR point cloud projection onto a 2D range image grid based on discretized angular coordinates.
Figure 2: Illustration of geometric differences between the ideal sensor model (perfectly aligned beams from a common origin) and the real-world model (beams with spatial and angular offsets), highlighting the calibration parameters required for lossless projection.
Figure 3: General outline of ALICE-LRI for estimating LiDAR intrinsic parameters and generating lossless range images. Gray diamond-shaped nodes represent data inputs and outputs; rectangular nodes represent processing steps. Solid lines denote execution flow, while dashed lines indicate data flow. The method first estimates sensor parameters from the input point cloud, then uses these parameters to generate range images from subsequent point clouds.
Figure 4: Flowchart of the vertical parameter estimation algorithm (Sec. 4.1 ). The algorithm iteratively identifies candidate scanline parameters using the Hough Transform, selects points consistent with these parameters, fits the scanline to refine parameter values, and resolves conflicts with previously accepted scanlines. The process continues until all points are assigned to scanlines or the search space is exhausted. Key processing steps are annotated with references to their corresponding subsections.
Figure 5: Comparison of basic Hough Transform strategy versus vote-for-discontinuities strategy. The basic strategy (a) misses the intersection peak due to vote fragmentation across cells, while the discontinuities strategy (b) successfully detects it by incrementing intermediate cells when votes skip multiple bins due to steep gradients in the parameter mapping.
Figure 6: Flowchart of the scanline fitting via weighted least squares procedure (Sec. 4.1.3 ). The algorithm performs iterative WLS fitting to refine vertical parameters (φ^(l),o^y(l)) and compute point memberships. A heuristic fallback is employed to handle cases where the fitting process fails to converge.
Figure 7: Flowchart of the scanline conflict resolution process (Sec. 4.1.4 ). The algorithm checks for intersections between the current scanline candidate and previously accepted scanlines. If intersections exist, uncertainty scores are compared to determine which scanlines to accept or reject. Rejected scanlines are stored for potential recovery if their conflicting counterparts are later invalidated.
Figure 8: Flowchart of the horizontal parameter estimation algorithm (Sec. 4.2 ). The algorithm performs an exhaustive search over candidate horizontal resolutions H^(l) for each scanline, computing horizontal offset o^x(l) and azimuthal offset θ^off(l) , and selects the parameters that minimize the loss function. For scanlines with insufficient points, a heuristic fallback reuses parameters from previously processed beams.
Symbol
Value
Reference
Oy
0.5m
Sec. 4.1.1
Δo~y
1\times10−3m
Sec. 4.1.1
Φ
π/2
Sec. 4.1.1
Δφ~′
1\times10−4rad
Sec. 4.1.1
ϵmin
1\times10−6m
Sec. 4.1.2
NWLS
10 iterations
Sec. 4.1.3
Table 1: Main numerical settings and thresholds used in the ALICE-LRI implementation.
Table 2: Characteristics of the datasets used for evaluation, showing sensor specifications and data distribution.
Dataset
Subset
Scanlines Count
# Samples
# Incorrect
OA (%)
mP (%)
mR (%)
mF1 (%)
wP (%)
wR (%)
wF1 (%)
KITTI
All
47,885
0
100.00
100.00
100.00
100.00
100.00
100.00
100.00
n(l)≥64
47,543
0
100.00
100.00
100.00
100.00
100.00
100.00
100.00
DurLAR
All
145,916
130
99.91
92.98
93.67
93.31
99.92
99.91
99.92
n(l)≥64
130,757
0
100.00
100.00
100.00
100.00
100.00
100.00
100.00
Table 3: Scanlines count estimation metrics on KITTI and DurLAR datasets. For each subset (all point clouds and those with n(l)≥64 ), the number of point clouds (# Samples), incorrect predictions (# Incorrect), overall accuracy (OA), macro and weighted precision (mP, wP), recall (mR, wR), and F1-score (mF1, wF1) are reported.
Dataset
Subset
Horizontal Resolution
# Samples
# Incorrect
OA (%)
KITTI
All
3,064,085
84
99.99
n(l)≥64
3,042,253
4
99.99
DurLAR
All
18,637,656
237
99.99
n(l)≥64
16,736,809
0
100.00
Table 4: Horizontal resolution estimation accuracy on KITTI and DurLAR. We report the total number of scanlines (# Samples), incorrect predictions (# Incorrect), and overall accuracy (OA) for all point clouds and for those with n(l)≥64 .
Dataset
Subset
Vert. Angle φ(l) (deg)
Vert. Offset oy(l) (mm)
Horiz. Offset ox(l) (mm)
Azim. Offset θoff(l) (deg)
MAX
MAE
MAX
MAE
MAX
MAE
MAX
MAE
KITTI
All
0.123479
0.000413
137.338350
0.057363
202.204798
0.040357
0.081500
0.000872
n(l)≥64
0.049864
0.000412
4.006125
0.056994
19.806281
0.038512
0.081500
0.000870
DurLAR
All
3.490045
0.000033
323.204766
0.005478
17.107649
0.003304
0.174496
0.000032
n(l)≥64
0.012686
0.000008
0.102721
0.000780
0.010094
0.002778
0.000117
0.000028
Table 5: Per-beam parameter estimation errors (MAE and MAX) for vertical angles, vertical and horizontal offsets, and azimuthal offsets on KITTI and DurLAR. Results are shown for all point clouds and for those with n(l)≥64 .
Algorithm Components
KITTI
DurLAR
Hough
Conflict
Vertical
Horizontal
# Incorrect
# Incorrect
# Incorrect
# Incorrect
Continuity
Resolution
Heuristics
Heuristics
Scanlines Count
Resolutions
Scanlines Count
Resolutions
E0
✓
✓
✓
✓
0 (0)
84 (4)
130 (0)
237 (0)
E1
✓
✓
✓
0 (0)
84 (4)
145 (0)
260 (0)
E2
✓
✓
✓
2 (0)
86 (4)
2110 (24)
3263 (154)
E3
✓
✓
✓
136 (3)
414 (7)
4137 (0)
13741 (0)
Table 6: Ablation study results showing the impact of different algorithm components on estimation accuracy. For each configuration, we report the number of incorrect scanline counts and resolution errors across KITTI and DurLAR datasets. Values not in parentheses represent all point clouds, while values in parentheses represent point clouds with n(l)≥64 . The full method (E0) includes all components, while subsequent configurations systematically disable specific features.
Dataset
Subset
Frames
Scanlines
# Total
# With Fallback
# Total
# With Fallback
KITTI
All
47,885
153 (0.320%)
3,064,085
469 (0.0153%)
n(l)≥64
47,543
3 (0.0063%)
3,042,253
3 (0.0001%)
DurLAR
All
145,916
10,507 (7.20%)
18,637,433
143,378 (0.769%)
n(l)≥64
130,757
0 (0%)
16,736,809
0 (0%)
Table 7: Fallback-trigger frequency for the full method (E0). The table reports frames and scanlines for which either the vertical or horizontal fallback is used, considering all frames and the dense subset ( n(l)≥64 for every scanline).
Figure 9: Representative sparse DurLAR frame in the vertical weighted least squares fitting space. Points are plotted as (1/r,φ) and colored by scanline using a cyclic categorical palette. The full fitting space shows the global distribution of scanline evidence, while the zoomed panels compare the estimated vertical fits with the corresponding ground-truth scanline references. Gray lines are standard WLS estimates, orange lines are heuristic fallback estimates, the red line is an incorrectly estimated fit, and green lines are ground-truth references. Labels identify scanlines for reference.
Dataset
Method
CD (m)
PSNR (dB)
SE (%)
AVG
MAX
AVG
MIN
AVG
MAX
KITTI
PBEA ( 4000$\times$64 )
0.027419
0.051127
63.056212
45.769613
8.689915
17.652985
PBEA ( 8000$\times$128 )
0.012521
0.022010
72.296912
51.615480
0.518697
3.034373
PBEA ( 16,000$\times$256 )
0.006156
0.010768
79.090223
52.142493
0.126372
0.755860
PBEA ( 32,000$\times$512 )
0.003055
0.005281
85.929010
52.152183
0.034610
0.196997
PBEA ( 64,000$\times$1024 )
0.001520
0.002664
92.654247
61.124456
0.009202
0.091909
Table 8: Point cloud reconstruction quality metrics comparing ALICE-LRI against PBEA baseline at native sensor resolutions and progressively increased resolutions. Results are computed over the complete datasets without excluding frames. For each method and resolution, average (AVG) and extreme frame-level (MAX/MIN) values for Chamfer Distance (CD), Peak Signal-to-Noise Ratio (PSNR), and Sampling Error (SE) are reported.
Figure 10: Chamfer Distance of ALICE-LRI and PBEA across different range image resolutions. ALICE-LRI (solid lines) consistently achieves lower reconstruction error compared to PBEA (dashed lines). Results show frame-by-frame reconstruction quality, where lower values indicate better geometric fidelity.
Dataset
Method
SP2P (m)
AVG
MAX
KITTI
PBEA ( 4000$\times$64 )
0.015650
0.033118
PBEA ( 8000$\times$128 )
0.007421
0.016817
PBEA ( 16,000$\times$256 )
0.003654
0.008142
PBEA ( 32,000$\times$512 )
0.001824
0.004040
PBEA ( 64,000$\times$1024 )
0.000910
0.002041
Table 9: Symmetric point-to-plane error (SP2P) for ALICE-LRI and PBEA across the evaluated range image resolutions on KITTI and DurLAR. The metric is computed using local PCA normals estimated from a 12-point nearest-neighbor neighborhood. AVG and MAX are computed over frame-level mean absolute SP2P values without excluding frames. Lower values indicate better local geometric preservation.
Figure 11: Qualitative comparison of 3D point cloud reconstruction quality. The highlighted regions (red boxes) illustrate the reconstruction fidelity achieved by ALICE-LRI compared to the baseline PBEA approach. ALICE-LRI maintains structural integrity while the baseline introduces visible artifacts and geometric distortions.
Figure 12: Comparison of range image quality between baseline and ALICE-LRI methods. The baseline method (top) exhibits visible discontinuities and missing data regions, while ALICE-LRI (bottom) produces clean, artifact-free range images that accurately represent the sensor scanning pattern. The highlighted regions show the improved continuity and data preservation achieved by ALICE-LRI.
Dataset
Estimation (s)
Proj. (ms)
Unproj. (ms)
KITTI
31.3±0.8
9.3±0.3
3.8±0.1
DurLAR
41.3±2.8
20.0±2.8
4.8±0.6
Table 10: Runtime performance analysis showing mean execution times ( ± standard deviation) for parameter estimation, projection, and unprojection phases on the KITTI and DurLAR datasets. Parameter estimation times represent one-time offline estimation costs, while projection and unprojection times indicate per-frame processing requirements for real-time applications.
Error Threshold
Compression Ratio
CD (m)
PSNR (dB)
SE (%)
Original
Modified
Original
Modified
Original
Modified
Original
Modified
0.001
10.7899
10.3009
0.0293
0.0020
62.4274
93.8161
10.0882
0.0009
0.010
10.4894
10.1392
0.0293
0.0021
62.4270
93.5504
10.0881
0.0009
0.050
12.2672
11.8046
0.0315
0.0059
62.3669
81.2060
10.0877
0.0008
0.100
15.0507
14.9141
0.0362
0.0130
62.1628
74.8480
10.0873
0.0007
0.250
19.5483
21.3198
0.0474
0.0301
61.4164
68.2742
10.0869
0.0005
Table 11: Point cloud compression performance comparison between the original RTST algorithm and the ALICE-LRI modified version. Results show Compression Ratio, Chamfer Distance (CD), Peak Signal-to-Noise Ratio (PSNR), and Sampling Error (SE) across different error thresholds. Bold values indicate superior performance for each metric.
Figure 13: Compression Ratio versus Chamfer Distance trade-off analysis. Each point represents one execution of the compression algorithm. Blue circles denote the original method and red triangles denote the ALICE-LRI modified method. The blue solid and red dashed lines show the corresponding linear regression fits, indicating that the ALICE-LRI modified method achieves higher compression ratios on average for equivalent reconstruction quality.
Error Threshold
Encoding Time (ms)
Decoding Time (ms)
Original
Modified
Overhead
Original
Modified
Overhead
0.001
20.79
23.24
2.45
3.84
6.70
2.86
0.010
19.71
26.00
6.29
4.08
6.78
2.70
0.050
22.57
28.70
6.13
4.85
7.34
2.49
0.100
30.25
37.67
7.43
4.05
7.55
3.51
0.250
43.27
49.89
6.63
4.47
7.52
3.05
Table 12: Mean runtime per frame for the original RTST algorithm and the ALICE-LRI modified version across different error thresholds. Execution times are reported in milliseconds. The minimal overhead introduced by ALICE-LRI ensures real-time viability while providing substantial quality improvements.
We present LV-Calib, a calibration framework for LiDAR-camera extrinsic estimation and LiDAR boundary-response calibration using a printable planar target. The target serves as a shared observation carrier: visual fiducials provide indexed image measurements, while circular reflectivity boundaries provide LiDAR-observable structural feature points. Instead of directly fitting boundary points as ideal geometric contours, LV-Calib automatically crops background points, estimates the target plane, and iteratively refines accurate LiDAR-side 3-D feature points from intensity and geometric constraints. The refinement explicitly handles the broadened and distorted transition band induced by finite beam footprint and mixed-intensity returns around black-white reflectivity discontinuities. Given these refined LiDAR features, we formulate a weighted reprojection-consistent extrinsic optimization with LiDAR feature alignment, where image observations are kept in the reprojection domain and LiDAR feature residuals are weighted by refinement confidence. Finally, using the estimated extrinsic and the extracted transition band, LV-Calib calibrates the LiDAR boundary response by estimating pitch-yaw-range residual statistics of boundary-overlap samples. Experiments on printed-board calibration data demonstrate sub-pixel reprojection accuracy, millimeter-level LiDAR feature consistency, and improved odometry performance. Code and calibration data will be released for reproducible evaluation.
Sheng Hong
Pen-Tung Sah Institute of Micro-Nano Science and Technology, Xiamen University, Xiamen 361102, China
Target-based LiDAR-camera extrinsic calibration is a prerequisite for multi-sensor fusion in robotics. However, in the widely adopted four-hole pipeline, \rt{calibration accuracy is limited by hole-center extraction in the LiDAR side, where sparse angular coverage and mixed-pixel returns displace the estimated centers}. This paper presents P2Calib, which exploits pattern priors, geometric constraints specified by the CAD model of the target board, to improve calibration accuracy. First, we incorporate the known hole radius as a fitting constraint to prevent center estimates from degrading under sparse angular coverage. Building on the improved hole estimates, we further enforce the rigid rectangular layout of the four holes as a global consistency constraint to correct residual errors across holes. Both priors are integrated into an \rt{interactive tool that runs the pipeline from target detection to the final extrinsic}. Experiments on simulated and real datasets show that P2Calib reduces the joint registration residual by 90% and 82% and the held-out reprojection error by 96% and 77% over the baseline. \rt{We will release the code and data\codelink{} to facilitate future research.
Xiangcheng Hu
Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong SAR, China
Accurate camera-LiDAR calibration is an essential prerequisite for robust multi-modal perception in robotics. Recent target-less methods based on learned pixel-point correspondences achieve high extrinsic calibration accuracy and generalize across environments. However, they operate on rectified images with known camera intrinsics, restricting their applicability to systems whose cameras have already been individually calibrated. We present a fully automatic, target-less framework that jointly estimates camera intrinsics, radial-tangential distortion, and camera-LiDAR extrinsics from raw images and LiDAR scans. Our method generalizes deep correspondence-based calibration from rectified images to raw cameras through intrinsic- and distortion-aware correspondence learning via synthetic augmentations. During calibration, we iteratively couple correspondence prediction and nonlinear optimization to jointly refine the camera parameters and relative sensor transformation. We evaluate our approach on an unseen camera-LiDAR configuration from the KITTI dataset and on three real robotic platforms in a zero-shot manner. The results demonstrate that joint calibration with our generalized pixel-point correspondence network improves extrinsic accuracy over correspondence-based methods using pre-calibrated cameras, while additionally recovering accurate intrinsics.
Simon Bultmann, Daniele Cattaneo, Abhinav Valada
Department of Computer Science, University of Freiburg, Germany