SLICE
SLICE

Scalable LiDAR Image Coding
with Embedded Refinement

An efficient neural codec for LiDAR range and intensity sequences.

Kang You1,†Jiahao Zhu1,†Dandan Ding2M. Salman Asif3Zhan Ma1,*
1 Nanjing University2 Hangzhou Normal University3 University of California, Riverside

† Equal contribution  ·  * Corresponding author

Watch demosExplore results Read abstract
63–76%Less lossless geometry bitrate

vs. G-PCC octree on WOD & SemanticKITTI

~3 bits / pointNear-uncompressed perception

Bitrate for detection & segmentation

10 FPSLossless edge-to-host streaming

Range + intensity · Jetson Thor → RTX 5070

01SLICE in the field

See the scene. Send fewer bits.

Indoor and outdoor recordings from the deployed system. Switch between the available scenes and transmission rates.

Field recordings · ROS2
Indoor · 400 kbpsOutdoor · 550 kbpsOutdoor · 270 kbpsSLAM · IndoorSLAM · Outdoor

Indoor · 400 kbps
Ouster OS1 measurements, encoded on the edge and reconstructed on the receiver.

Open recording

02The idea

Record completely.
Transmit progressively.

Autonomous platforms need to keep up with the sensor while managing limited storage and communication bandwidth. SLICE addresses both needs with embedded lossy-to-lossless coding.

Read the full abstract

LiDAR compression for autonomous platforms must support storage-efficient onboard caching and bandwidth-efficient scene transmission without falling behind sensor acquisition. We present SLICE, a scalable neural codec for range and intensity sequences that supports both applications through embedded lossy-to-lossless coding. A reversible multiscale representation interleaves spatial and amplitude decompositions for range images, while intensity is progressively refined on the decoded range grid. Cross-scale conditional models exploit lower-scale reconstructions and temporal references for both modalities, with additional range guidance for intensity. Each modality-specific model supports multiple bitrates and both intra- and inter-frame coding, while one-step restoration enables efficient reconstruction from truncated bitstreams. On WOD and SemanticKITTI, SLICE reduces lossless geometry bitrate by 63.01–76.20% relative to G-PCC octree coding, substantially lowering the storage required for exact recovery of the coded measurements. At approximately 3 bits per point for range coding, it also preserves near-uncompressed detection and segmentation accuracy. A quantized implementation sustains lossless range-and-intensity streaming at 10 FPS in a ROS2 pipeline from a Jetson Thor encoder to an RTX 5070 laptop decoder. These results demonstrate a practical codec for compact onboard recording and scalable remote reconstruction of LiDAR scenes.

SLICE overview: supported coding modes, spatial and amplitude refinement, and rate-distortion efficiency versus encoding speed.
SLICE at a glance. Range and intensity, intra- and inter-frame coding, and progressive lossy-to-lossless reconstruction. The throughput comparisons shown here use an RTX 5080 GPU.

03How it works

A coarse-to-fine path to lossless.

SLICE codes refinements between reversible scales. Each modality has its own model, shared across bitrates and intra- / inter-frame modes.

* Most LiDAR sensors natively represent range measurements as unsigned integers; lossless coding here refers to the exact reconstruction of these integer values.

01 / REPRESENT

Build a multiscale hierarchy

Interleave spatial sampling and amplitude quantization for range images. Refine intensity amplitude on the spatial grid established by decoded range.

02 / CONDITION

Predict each refinement

Use decoded lower-scale information and, for inter-frame coding, the previous frame’s same-scale context. Decoded range also guides intensity coding.

03 / RECONSTRUCT

Stop early or recover exactly

One-step restoration predicts missing amplitude detail after truncation. Decoding every refinement bypasses restoration and recovers the coded measurements exactly.

SLICE framework with range and intensity acquisition, reversible decomposition operators, multiscale hierarchies, and conditional coding branches.
The coding framework. Lower-scale context extraction, conditional context fusion, refinement residual coding, and one-step restoration are shared building blocks. Valid-return maps identify the samples to code during spatial range refinement.
Explore progressive range and intensity reconstruction

Horizontal refinements improve geometry; vertical refinements improve intensity at the selected geometry level. Truncated reconstructions retain the spatial resolution of the decoded range scale.

Multiscale reconstruction examples with geometry refinement across columns and intensity refinement down rows, annotated with bitrate, PSNR and decoding time.
Scalable operating points from the paper, with bitrate, reconstruction quality, and decoding latency for each level.

04Experimental results

Fewer bits. Useful scenes.

Evaluated on the Waymo Open Dataset (WOD) and SemanticKITTI, from signal fidelity to 3D detection and semantic segmentation.

Scalable rate–distortion performance

WOD + SEMANTICKITTI
Dataset and evaluation protocol

Waymo Open Dataset

Native 64 × 2,650 first-return range and intensity images. Range is capped at 75 m and quantized to 1 mm; intensity is clipped to [0, 1] and mapped to 8-bit values.

SemanticKITTI

Spherical projection produces 64 × 2,048 pseudo images. Range is capped at 82 m at 1 mm precision; intensity is bounded to [0, 99]. Sequence 08 is reserved for validation.

Rate-distortion comparisons across D1 PSNR, D2 PSNR, Chamfer distance, and reflectance PSNR on WOD and SemanticKITTI.
Rate–distortion curves. WOD in the top row; SemanticKITTI in the bottom row. SLICE improves coding efficiency across the evaluated geometry and intensity operating points.
Full geometry and intensity benchmark tables
Full geometry benchmark comparing G-PCC, RENO, SCP-EHEM, PACE, JPEG XL, RangeCM, BitRIC, and SLICE.
Geometry compression: lossy BD-rate and lossless bits per point.
Full intensity benchmark comparing G-PCC RAHT, JPEG XL, RangeCM, and SLICE.
Intensity compression: lossy BD-rate and lossless bits per point.
Compare reconstructed geometry and intensity
Point-cloud geometry and intensity error maps comparing SLICE, RangeCM, JPEG XL, and G-PCC at two rate points.
Geometry error is measured by point-to-point distance. Attribute error is the absolute intensity difference of nearest neighbors, clipped at 20 for visualization.

Preserve the information perception needs

CenterPoint detection on WOD and SparseUNet segmentation on SemanticKITTI evaluate the reconstructed point clouds. Perception models are trained on uncompressed inputs and frozen during testing.

CenterPoint object detection accuracy on WOD versus bitrate for vehicle, pedestrian, cyclist, and overall mean AP.
3D object detection · WOD · CenterPoint. Geometry-only XYZ inputs in the top row; geometry-plus-intensity XYZI inputs in the bottom row. Dashed lines mark uncompressed performance.
SparseUNet semantic segmentation on SemanticKITTI versus bitrate for class IoU, mean accuracy, overall accuracy, and mIoU.
Semantic segmentation · SemanticKITTI · SparseUNet. Geometry-only XYZ inputs in the top row; geometry-plus-intensity XYZI inputs in the bottom row.

Efficient at every operating point.

Desktop benchmark on an AMD Ryzen 9950X3D and a single RTX 5080. Lossy timings are averaged across evaluated bitrates.

Runtime and model complexity across all baselines
Full lossy and lossless runtime and model-parameter comparison across rule-based and learned codecs.
Average encoding and decoding time, with model parameter counts. These desktop measurements are separate from the deployed edge results below.

05Real-world deployment

On the robot.
Across the link.

SLICE-Q, the quantized implementation, runs in a ROS2 pipeline between a Jetson Thor edge encoder and an RTX 5070 laptop decoder.

AgileX mobile robot with Ouster OS1 LiDAR and NVIDIA Jetson Thor, connected to an RTX 5070 laptop.
The deployed platform. An AgileX UGV carries the LiDAR and edge compute node; the laptop receives and visualizes the reconstructed scene.
Sensor
Ouster OS1 · 1,024 × 64 · 10 Hz
Edge encoder
NVIDIA Jetson Thor
Remote decoder
NVIDIA GeForce RTX 5070 laptop
Transport & integration
Wi-Fi 6 · ROS2 · SLICE-Q
  1. 01 / ACQUIRERange + intensityNative organized images
  2. 02 / ENCODEJetson ThorDirect image compression
  3. 03 / TRANSMITEmbedded bitstreamWi-Fi 6 link
  4. 04 / RECONSTRUCTRTX 5070Decode, then project to 3D
BOTH MODALITIES · LOSSLESS

Keeping pace
with acquisition.

The dual-lossless endpoint supports 10 FPS streaming, matching the configured OS1 acquisition rate.

94 msEDGE ENCODING / FRAME
82 msHOST DECODING / FRAME

These are combined range-and-intensity coding times on each device; they do not include network transport latency.

Per-scale range and intensity encoding latency on Jetson Thor and decoding latency on RTX 5070, with real-time thresholds.
Latency across refinement scales. Range scale 16 and intensity scale 8 are lossless. Intensity timings use full-resolution, lossless range as context.

Scene structure at 290 kbps

Watch recordings ↑

On the OS1 scan shown below, SLICE retains ground contours and local object shapes at 290 kbps, compared with JPEG XL at 311 kbps and G-PCC at 307 kbps.

Live OS1 scan: original data and reconstructions from SLICE at 290 kbps, JPEG XL at 311 kbps, and G-PCC at 307 kbps, with enlarged regions.
Reconstruction on live LiDAR data. Color represents intensity, from cyan (low) to purple (high). Insets highlight ground contours and object detail.

06People behind SLICE

Our Team.

Nanjing University

Kang You†

Jiahao Zhu†

Zhan Ma*

Hangzhou Normal University

Dandan Ding

University of California, Riverside

M. Salman Asif

† Equal contribution.  ·  * Corresponding author.

Research figure