PaperlayerView exact revision on arXiv

Source-linked AI summary

ALICE-LRI: A General Method for Lossless Range Image Generation for Spinning LiDAR Sensors without Calibration Metadata

Samuel Soutullo, Miguel Yermo, David L. Vilariño, Óscar G. Lorenzo, José C. Cabaleiro, Francisco F. Rivera

arXiv:2510.20708v1cs.CVcs.RO

TL;DR

Conventional projection can lose information because factory-calibrated LiDAR geometry violates the spherical model. ALICE-LRI infers sensor intrinsics from point clouds without factory metadata, achieving zero point loss and lower geometric error than PBEA on KITTI. The authors identify motion-related extrinsic corrections as a scope boundary.

  • Problem

    Factory-calibrated beam geometry violates spherical projection assumptions, causing projection mismatches, pixel collisions, and quantization artifacts.

    Source: §1 · p. 1, §3.3 · p. 4

  • Method

    ALICE-LRI infers spinning-LiDAR intrinsic parameters directly from point clouds and uses them for lossless range image generation.

    Source: §1 · p. 2

  • Results

    ALICE-LRI achieved zero point loss and geometric accuracy orders of magnitude below sensor precision; PBEA had 0.027 m Chamfer Distance and 8.69% sampling losses on KITTI.

    Source: §8 · p. 23

  • Takeaways & Limitations

    The authors report real-time suitability and better compression efficiency without sacrificing point-cloud quality.

    Source: §8 · p. 23

    The approach focuses on intrinsic parameters, while motion-correction extrinsics can also make the spherical model deviate from the static sensor model.

    Source: §9 · p. 24

Abstract

from arXiv · show

3D LiDAR sensors are essential for autonomous navigation, environmental monitoring, and precision mapping in remote sensing applications. To efficiently process the massive point clouds generated by these sensors, LiDAR data is often projected into 2D range images that organize points by their angular positions and distances. While these range image representations enable efficient processing, conventional projection methods suffer from fundamental geometric inconsistencies that cause irreversible information loss, compromising high-fidelity applications. We present ALICE-LRI (Automatic LiDAR Intrinsic Calibration Estimation for Lossless Range Images), the first general, sensor-agnostic method that achieves lossless range image generation from spinning LiDAR point clouds without requiring manufacturer metadata or calibration files. Our algorithm automatically reverse-engineers the intrinsic geometry of any spinning LiDAR sensor by inferring critical parameters including laser beam configuration, angular distributions, and per-beam calibration corrections, enabling lossless projection and complete point cloud reconstruction with zero point loss. Comprehensive evaluation across the complete KITTI and DurLAR datasets demonstrates that ALICE-LRI achieves perfect point preservation, with zero points lost across all point clouds. Geometric accuracy is maintained well within sensor precision limits, establishing geometric losslessness with real-time performance. We also present a compression case study that validates substantial downstream benefits, demonstrating significant quality improvements in practical applications. This paradigm shift from approximate to lossless LiDAR projections opens new possibilities for high-precision remote sensing applications requiring complete geometric preservation.

1. Introduction

Conventional spherical projections can lose geometric information when real sensors’ factory calibration violates ideal geometric assumptions. ALICE-LRI estimates sensor intrinsics from point clouds to generate lossless range images without manufacturer metadata.

1. Introduction

  • Factory-calibrated beam positions and orientations violate the spherical projection model, causing projection mismatches, pixel collisions, and quantization artifacts.The ideal model assumes discrete vertical and horizontal angles, which permit exact reconstruction.Source: §1 · p. 1
  • Prior methods that model laser-specific offsets rely on manufacturer lookup tables or packet-level data, limiting their generality.The paper identifies lossless generation from calibrated point clouds without sensor metadata as an overlooked challenge.Source: §1 · p. 2
  • The work formulates lossless range image generation and analyzes how real-world sensor geometry breaks conventional projection methods.Its evaluation covers KITTI and DurLAR and includes a compression case study.Source: §1 · p. 2
  • ALICE-LRI is presented as the first method to generate lossless range images from calibrated spinning-LiDAR point clouds without manufacturer metadata or calibration files.It infers sensor intrinsics directly from raw data.Source: §1 · p. 2

2. Related Work

The related work covers range-image applications and LiDAR calibration, emphasizing that projection quality matters and existing calibration approaches generally estimate parameters from raw sensor data.

2. Related Work

  • The related-work review spans range-image applications and research on LiDAR sensor geometry and calibration.Real-world sensors exhibit geometric non-idealities from manufacturing tolerances and mechanical constraints.Source: §2 · p. 2, §2.2 · p. 3

2.1. Range Image Use Cases

  • Range images support segmentation, detection, upsampling, odometry and localization, and compression, but their utility depends on projection fidelity.Compression methods assume faithful range images; projection-related information loss can affect point-cloud reconstruction.Source: §2.1 · pp. 2–3

2.2. LiDAR Sensor Geometry and Calibration

  • Earlier methods refine beam orientations and offsets; later work evaluates calibration stability or models additional horizontal-angle and range offsets.The reviewed approaches use structured-environment features and sensor-specific calibration models.Source: §2.2 · p. 3
  • ALICE-LRI addresses the inverse problem of recovering applied calibration parameters from already-calibrated point clouds, including datasets without raw packets or calibration metadata.Prior calibration research is framed as a forward problem: estimating parameters from raw data to produce accurate point clouds.Source: §2.2 · p. 3

3. Problem Formulation

Section 3 formulates lossless range-image generation as a bijective mapping between 3D points and a discretized angular grid. It explains why ideal geometry permits exact reconstruction, while real-world beam geometry requires recovering sensor-specific parameters to restore that mapping.

3.1. Range Image Representation

  • A range image maps spherical coordinates (range and angles) onto a regular 2D grid, with each pixel encoding range along a discretized angular direction.The assumed grid has L laser-beam elevations and H uniformly sampled azimuth directions.Source: §3.1 · p. 3

3.2. Ideal Sensor Model

  • In the ideal model, beams share an origin and have fixed vertical angles, while rotation yields uniformly spaced azimuth samples, enabling exact point reconstruction.Each point is uniquely determined by its range and corresponding beam and azimuth angles.Source: §3.2 · p. 4

3.3. Real-World Sensor Model

  • Real-world beam offsets and deviations make angles continuous rather than discrete, causing pixel misalignment, collisions, and holes that prevent standard projection from being lossless.The model identifies spatial offsets, range deviations, and azimuthal offsets as geometric non-idealities.Source: §3.3 · pp. 4–5
  • The real-world model characterizes each beam with a vertical angle, spatial offsets, and beam-specific horizontal resolution, while allowing nonuniform resolutions for generality.The formulation omits forward-backward spatial corrections because they can be absorbed into horizontal and vertical offsets.Source: §3.3 · p. 4

3.4. Lossless Range Image Generation

  • Recovering unknown beam-specific parameters allows correction of observed angles to the discrete sets needed for bijective, lossless projection.The required parameter set includes per-beam vertical angles, offsets, horizontal angular corrections, and horizontal resolutions.Source: §3.4 · p. 5
  • For lossless projection, corrected angles must return to discrete beam and azimuth sets; the image dimensions depend on L and the least common multiple of the H(l) values.The required correction reverses calibration effects on the observed spherical angles.Source: §3.4 · p. 5

4. ALICE-LRI

ALICE-LRI estimates spinning LiDAR intrinsic parameters from point clouds without sensor metadata or scene knowledge, then uses them for lossless range-image projection. Estimation separates vertical and horizontal parameter recovery, and the recovered parameters can be reused for scans from the same device.

4. ALICE-LRI

  • ALICE-LRI estimates sensor parameters from a point cloud and uses them to generate lossless range images, without requiring sensor metadata or scene knowledge.The method has parameter-estimation and lossless-projection stages.Source: §4 · pp. 5–6
  • The recovered parameters need to be estimated once per device and can then be reused to generate range images from multiple point clouds captured by that hardware.The projection stage applies the estimated parameters to subsequent point clouds.Source: §4 · p. 6
  • Vertical parameter estimation determines the beam count, assigns points to beams, and estimates each beam's vertical angle and vertical offset.These estimates are part of the intrinsic parameters recovered from the point cloud.Source: §4 · p. 6
  • Horizontal parameter estimation uses the beam assignments to estimate each beam's horizontal angular resolution and horizontal offset.This follows the vertical estimation phase.Source: §4 · p. 6

4.1. Vertical Parameter Estimation: L, ϕ(l), and o(l)

This section describes ALICE-LRI’s iterative estimation of vertical scanlines and their beam-specific angles and offsets, combining candidate generation, point assignment, uncertainty-aware fitting, and global consistency checks. It also explains how geometric conflicts and fit uncertainty guide candidate resolution, while rejected candidates remain available for possible recovery.

4.1. Vertical Parameter Estimation: L, ϕ(l), and o(l)

  • The algorithm iteratively proposes scanline parameters, assigns consistent points, refines the parameters, and resolves conflicts until points are assigned or the search space is exhausted.It uses consensus-based steps to recover scanlines that best fit the observations under the vertical model.Source: §4.1 · p. 6

4.1.1. Candidate Scanline Parameter Identification via the Hough Transform

  • Hough Transform peaks identify candidate vertical angles and offsets, but noise, sparse data, and accumulator discretization can produce spurious or fragmented candidates.The method therefore uses Hough only for initial candidates, which are refined and validated in later stages.Source: §4.1.1 · pp. 7–8

4.1.2. Scanline Point Selection via Error Bounds

  • Point selection accounts for coordinate quantization by inferring its step from minimum nonzero coordinate differences and propagating the resulting coordinate uncertainty into angular error bounds.The quantization estimate assumes at least one axis contains two occupied adjacent levels; the angular bound then determines whether a point belongs to a candidate scanline.Source: §4.1.2 · pp. 8–9

4.1.3. Scanline Fitting via Weighted Least Squares

  • Weighted least squares fits scanline angle and offset after linearizing the model for offsets small relative to range, weighting each point by the inverse squared angular error bound.The method iteratively updates point membership and refits until assignments stabilize or 10 iterations are reached; a neighboring-beam offset heuristic is used if fitting fails to converge.Source: §4.1.3 · pp. 9–10

4.1.4. Scanline Conflict Resolution

  • Conflict resolution prevents multiple scanlines from claiming the same points or forming implausible intersections, comparing uncertainty scores and retaining rejected scanlines for possible recovery.The step supports backtracking from suboptimal local choices to improve global consistency.Source: §4.1.4 · p. 10, Figure 7 · p. 11
  • Scanline conflicts are checked both when points are claimed by multiple scanlines and when fitted scanlines or their confidence intervals overlap within the point cloud’s valid range.A candidate with no conflicts is accepted, its points are assigned, and its votes are removed from the Hough accumulator.Source: §4.1.4 · p. 11
  • A candidate with higher uncertainty than any conflicting scanline is rejected, while one more reliable than all conflicts is accepted and invalidates those scanlines.Rejected candidates and their rejection metadata are retained; candidates without remaining conflicts are reintroduced into the Hough accumulator, allowing recovery from earlier choices.Source: §4.1.4 · p. 11
  • When scanlines are rejected, a hash-based cell reset prevents redundant processing of identical scanlines.The reset procedure is the same one described for Section 4.1.3.Source: §4.1.4 · p. 11

4.2. Horizontal Parameter Estimation: H(l), o(l) x , and θ (l)

ALICE-LRI estimates horizontal resolution and offsets independently for each scanline, selecting parameters that best fit its observations. A heuristic fallback reuses information from previously processed beams when a scanline has too few points for reliable estimation.

4.2. Horizontal Parameter Estimation: H(l), o(l) x , and θ (l)

  • For each scanline, an exhaustive search selects the candidate horizontal resolution and offsets that minimize a normalized fitting loss.The loss compares observed azimuths with the ideal angles implied by each candidate resolution.Source: §4.2 · p. 11
  • When the candidate resolution is correct, azimuth discrepancy varies piecewise linearly with rotation angle, with slope determined by the horizontal offset.The method segments these pieces, estimates local slopes, and combines them using a point-count-weighted median.Source: §4.2 · p. 12
  • Offset estimation segments azimuth discrepancies using thresholds tied to the resolution period and an empirically chosen discrepancy threshold.The angular threshold is one quarter of the piecewise function’s period; the discrepancy threshold is 10^-2 m^-1 and is validated for typical 1–100 m LiDAR ranges.Source: §4.2 · p. 12

4.2.2. Heuristic Fallback

  • For scanlines with insufficient points, the fallback evaluates candidate parameters using information from previously estimated beams and selects the pair with minimum heuristic loss.It estimates the azimuthal offset from the mean deviation between corrected and ideal angles, with resolution-normalized loss values.Source: §4.2.2 · p. 13

4.3. Range Image Generation and Point Cloud Reconstruction

With estimated intrinsics, ALICE-LRI maps points to a range image and back-projects that image to reconstruct the original point cloud. The transformation is lossless up to floating-point precision under the assumption that the intrinsics are correct.

4.3. Range Image Generation and Point Cloud Reconstruction

  • Under correctly estimated intrinsics, range-image generation and reconstruction preserve the point cloud without information loss, up to floating-point precision.This bidirectional transformation is the stated objective of the approach.Source: §4.3 · p. 13, §4.3.2 · p. 14

4.3.1. Point Cloud to Range Image

  • Each point is assigned to the beam whose vertical parameters best explain its elevation, then its range is stored at the corresponding image pixel.The image height is the number of beams, while its width is the least common multiple of beam resolutions to provide a uniform grid.Source: §4.3.1 · p. 14

4.3.2. Range Image to Point Cloud

  • For reconstruction, the pixel row identifies the beam, and the observed angles are recovered before converting the range and angles to Cartesian coordinates.The pixel’s beam index is L − v_i − 1; reconstruction adds back angular residuals.Source: §4.3.2 · p. 14

5. Materials and Methods

The evaluation uses the complete KITTI and DurLAR datasets, which represent distinct sensor configurations and data challenges. It assesses parameter estimation, reconstruction quality, and runtime, using complementary metrics for point preservation and geometric fidelity.

5.1. Datasets

  • KITTI and DurLAR provide contrasting evaluation settings: substantial geometric distortions in KITTI and sparse or missing scanlines at extreme viewing angles in DurLAR.The datasets contain Velodyne HDL-64E and Ouster OS1-128 scans, respectively.Source: §5.1 · p. 14

5.2. Experimental Setup

  • Experiments process all 193,801 point clouds across both datasets, with runtime measured separately on a local workstation.The full-dataset results and parameters were stored in SQLite databases for reproducibility and downstream analysis.Source: §5.2 · p. 14

5.3. Evaluation Metrics

  • The evaluation covers sensor-parameter accuracy, reconstructed point-cloud quality, and runtime performance, with a compression case study demonstrating practical benefits.Parameter estimation uses classification metrics and MAE; reconstruction uses geometric similarity metrics.Source: §5.3 · pp. 14–15
  • Sampling Error measures relative point-count difference, while the adopted non-squared, averaged Chamfer Distance expresses average displacement in meters.PSNR is also used, with maximum representable ranges of 120 m for KITTI and 170 m for DurLAR.Source: §5.3 · p. 15

6. Evaluation

ALICE-LRI is evaluated on KITTI and DurLAR for parameter estimation, reconstruction quality, and runtime, using the same methodology across both datasets. Its estimation results are strongest on adequately populated scanlines, while heuristics provide fallbacks for sparse data.

6. Evaluation

  • The evaluation combines quantitative metrics and visualizations to assess parameter estimation accuracy, reconstruction fidelity, and computational performance on KITTI and DurLAR.The code and experiment configurations are provided publicly for reproducibility.Source: §6 · p. 15

6.1. Parameter Estimation

  • ALICE-LRI estimates scanline count, vertical angles, spatial offsets, horizontal resolution, and azimuthal offsets against manufacturer and calibration references.The evaluation uses frame-specific scanline-count references because beams may occasionally fail to return points; the same methodology is applied to both datasets without dataset-specific customization.Source: §6.1 · pp. 15–16, §6 · p. 15

6.1.1. Scanlines Count

  • 100% scanline-count accuracy on KITTI and 99.91% on DurLAR across all point clouds; both reach 100% on scans with at least 64 points per scanline.The gap between macro and weighted DurLAR metrics reflects the difficulty of low-density scans with missing scanlines, not a method limitation.Source: §6.1.1 · p. 16

6.1.2. Per-Beam Parameters

  • For scans with at least 64 points per scanline, vertical-angle MAE is 4.12 · 10^-4° on KITTI and 8 · 10^-6° on DurLAR, with similarly accurate offset estimates.On KITTI, the maximum horizontal offset error falls from 19.8 mm to 6.3 mm when only correct resolution estimates are considered; sparse scans and incorrect point-to-scanline assignments are primary error sources.Source: §6.1.2 · pp. 16–17

6.1.3. Ablation Study

  • The complete method performs best overall, with disabling components increasing errors, especially on sparse data and the more challenging DurLAR dataset.Removing both heuristics has little effect on dense scans, where performance remains nearly perfect.Source: §6.1.3 · p. 18

6.2. Range Image Error Analysis

This section tests whether inferred sensor parameters improve range-image reconstruction against PBEA, using quantitative metrics and visual inspection. Across complete KITTI and DurLAR datasets, ALICE-LRI preserves all points and keeps geometric error below sensor noise.

6.2.1. Quantitative Analysis

  • At native resolution, ALICE-LRI has 0% sampling error on both datasets, versus 8.69% for PBEA on KITTI and 3.52% on DurLAR.Average Chamfer Distance is 3.80 · 10^-4 m for ALICE-LRI versus 0.027 m for PBEA on KITTI, and 1 · 10^-6 m versus 0.024 m on DurLAR.Source: §6.2.1 · p. 19
  • Even after PBEA resolution is scaled 32-fold in each dimension, it retains non-zero sampling error while ALICE-LRI’s Chamfer Distance remains below each sensor’s intrinsic noise level.The increased PBEA grids contain 1024 times as many pixels as the native-resolution grids.Source: §6.2.1 · p. 19
  • Across representative frames, ALICE-LRI maintains sub-millimeter Chamfer Distance while PBEA has higher reconstruction error, including at increased resolutions.The frame-level analysis reports that ALICE-LRI’s advantage is consistent across diverse scanning conditions and scene contents, rather than driven by outliers.Source: §6.2.1 · p. 19

6.2.2. Qualitative Analysis

  • Visual comparisons show that PBEA introduces point-cloud distortions and range-image gaps, whereas ALICE-LRI preserves structure and produces smooth, continuous scans.The paper notes that cleaner range images may benefit segmentation, object detection, and scene understanding, but does not evaluate those applications.Source: §6.2.2 · p. 19

6.3. Runtime Performance

The section assesses reconstruction quality and per-frame runtime on KITTI and DurLAR. ALICE-LRI’s combined projection and unprojection fit within the 10 Hz frame budget, leaving computational headroom.

6.3. Runtime Performance

  • Projection and unprojection take 9.3 ms and 3.8 ms per KITTI frame, and 20.0 ms and 4.8 ms per DurLAR frame.The paper attributes DurLAR’s longer processing times to its higher scanline count.Source: §6.3 · pp. 19, 22
  • ALICE-LRI produces lower reconstruction error than PBEA across range-image resolutions, while preserving point-cloud structure and avoiding visible artifacts.The accompanying comparisons also show continuous range-image patterns without missing data regions.Source: Figure 9 · p. 20, Figure 10 · p. 21, Figure 11 · p. 21
  • 13.1 ms on KITTI and 24.8 ms on DurLAR are required for combined projection and unprojection, equal to 13.1% and 24.8% of the 100 ms frame period.Both sensors operate at 10 Hz, leaving substantial time for downstream processing.Source: §6.3 · p. 22

7. Application

A KITTI case study replaces only RTST’s range-image projection and unprojection with ALICE-LRI. It reports better point-cloud quality at low error thresholds and higher average compression ratios for equivalent Chamfer Distance, with modest runtime overhead.

7. Application

  • The case study compares original RTST with a version that changes only range-image projection and unprojection, and is limited to KITTI.The publicly released RTST code is tailored to KITTI, and the authors leave its other components unchanged.Source: §7 · p. 22
  • At error threshold 0.001, ALICE-LRI reduces Chamfer Distance from 0.0293 m to 0.0020 m and raises PSNR from 62.43 dB to 93.82 dB.Sampling Error falls from 10.09% to 0.0009% at that threshold; the quality gap narrows for Chamfer Distance and PSNR as compression artifacts increasingly dominate.Source: §7 · p. 22
  • For a given Chamfer Distance, the ALICE-LRI-modified RTST achieves higher compression ratios on average than the original method.Compression ratios show mixed results across thresholds, but the paper reports better quality at equivalent compression ratios, allowing more aggressive compression settings without significant quality degradation.Source: §7 · p. 22, Figure 12 · p. 23

7.1. Runtime Performance Analysis

  • ALICE-LRI adds 2–8 ms to encoding and 2–3 ms to decoding, while total execution remains within available real-time headroom.The paper reports this overhead across error thresholds, including the highest thresholds where compression takes longest.Source: §7.1 · p. 23

8. Conclusion

ALICE-LRI estimates spinning LiDAR geometry to produce lossless range images without sensor-specific metadata. On KITTI and DurLAR, it preserves all points, maintains geometric accuracy below sensor precision, and supports real-time processing and improved compression at equivalent reconstruction fidelity.

8. Conclusion

  • 0 points were lost; geometric accuracy was orders of magnitude below sensor precision, while PBEA had 0.027 m Chamfer Distance and 8.69% sampling losses on KITTI.The evaluation covered KITTI and DurLAR; the cited baseline values are specifically reported for KITTI.Source: §8 · p. 23
  • Projection and unprojection complete in milliseconds; ALICE-LRI adds minimal overhead to RTST while remaining real-time, and improves compression at equivalent reconstruction fidelity.The compression comparison reports higher ratios at equivalent reconstruction fidelity.Source: §8 · p. 23, Table 9 · p. 24
  • The sensor-agnostic design requires neither manual tuning nor manufacturer metadata, and the open-source implementation supports adoption across diverse LiDAR platforms.The paper presents this as a deployment benefit for applications requiring high-precision geometric modeling.Source: §8 · p. 23
  • Removing manufacturer-metadata requirements lowers a barrier to deploying LiDAR-based systems across diverse sensor platforms and application domains.The paper positions ALICE-LRI as a foundation for more reliable LiDAR processing in autonomous vehicles, robotics, and mapping.Source: §8 · pp. 23–24

9. Future Work

The paper identifies several directions for extending ALICE-LRI: modeling extrinsic motion effects, evaluating downstream perception tasks and sensor-aware data generation, and accelerating its single-threaded implementation.

9. Future Work

  • The current method focuses on intrinsic parameters, while motion-corrected datasets can include extrinsic ego-motion effects; future work could estimate both together.The paper cites KITTI odometry as an example and notes that uncorrected raw data or known extrinsic parameters are often available in practice.Source: §9 · p. 24
  • Future evaluations could test effects on semantic segmentation, object detection, and scene understanding; inferred geometry could also support sensor-aware upsampling for sparse automotive LiDAR.The authors also propose generating training datasets that reflect sensor-specific properties for machine learning applications.Source: §9 · p. 24
  • The current implementation is single-threaded despite real-time viability; multithreading or GPU acceleration could reduce execution times and support high-throughput deployment.Source: §9 · p. 24
Loading 2510.20708v1…