LIORNet

What It Is

LIORNet is a 2026 self-supervised LiDAR snow-removal framework for autonomous driving under adverse weather.
It targets spurious LiDAR returns from falling snow and is explicitly motivated by degradation in snow, rain, and fog, but its demonstrated filtering task is snow removal.
It combines learned segmentation with physical and statistical priors instead of relying only on fixed distance or intensity thresholds.
It is not a general object detector, localization system, or full adverse-weather perception stack.
It should be read beside Radar-LiDAR Fusion in Adverse Weather, Production Perception Systems, and learned snow filters such as LiSnowNet and SLiDE.

Convert each LiDAR sweep into a structured range image with range, intensity, and reflectivity channels.
Train a U-Net++ encoder-decoder to predict a point-wise snow mask from that range/intensity/reflectivity range image.
Use residual blocks, nested skip connections, bilinear upsampling, and deep supervision to stabilize dense mask learning on sparse, noisy data.
Avoid manual snow labels by generating pseudo-labels from physics and rules: range-dependent intensity thresholds, reflectivity consistency, spatial sparsity, edge conditions, and sensing-range constraints.
Produce a sigmoid snow mask, then map it back one-to-one to the original LiDAR points so the output point cloud preserves the input ordering and geometry for non-snow points.
Apply physics/rule post-processing after the learned mask to correct false positives using LiDAR height, plausible snow sensing distance, ground conditions, and range constraints.

Input: one LiDAR sweep with per-point range, intensity, and reflectivity values that can be projected into a range image.
Input during training: unlabeled snowy point clouds plus pseudo-labels generated from LIORNet's physical and statistical rules.
Output: a binary or probabilistic snow mask over the range image.
Output: a denoised point cloud formed by removing points classified as snow.
Intermediate output: supervised heads at multiple decoder depths when deep supervision is enabled.
Non-output: it does not classify objects, estimate motion, correct ego pose, restore occluded geometry, or produce radar corroboration.

Project the raw point cloud to a spherical range-image representation with channels for range, intensity, and reflectivity.
Generate pseudo-labels by sequentially applying intensity, reflectivity, edge, sparsity, and range conditions to identify likely snow points.
Feed the three-channel range image into a U-Net++ style encoder-decoder.
Use residual blocks made from convolution, batch normalization, and ReLU layers.
Use nested skip paths to fuse shallow spatial detail with deeper semantic context.
Use deep supervision to train multiple decoder outputs instead of relying only on the final mask.
Apply a sigmoid activation to obtain the snow probability mask.
Reproject the mask back to the original point cloud with a one-to-one point mapping.
Apply physics/rule post-processing to suppress implausible snow predictions and preserve environmental structures.

The paper trains on WADS and evaluates on WADS, CADC, and self-collected snow data from South Korea, Sweden, and Denmark.
Reported baselines include DROR, LIOR, D-LIOR, DSOR, DDIOR, and LiSnowNet.
The best reported configuration combines U-Net++, deep supervision, and post-processing.
The paper reports strong recall and a 43.5 Hz runtime for the U-Net++ with post-processing configuration on WADS, which is above common 10-20 Hz spinning-LiDAR frame rates.
The authors report that post-processing reduces false-positive snow classifications, improving the precision/recall balance.
Evaluation is primarily about snow/noise filtering and structural preservation, not downstream safety-case performance.

Keeps useful hand-engineered physics in the loop while adding a learned mask for more adaptive snow separation.
Uses range, intensity, and reflectivity together instead of depending on a single threshold.
Avoids manual snow annotation by using pseudo-labels, which matters because snow points are small, dense, and transient.
The one-to-one point mapping makes integration easier for pipelines that expect the original point ordering or per-point metadata.
U-Net++ nested skips and deep supervision are a reasonable fit for thin structures, curbs, road boundaries, poles, and walls in range-image form.
Post-processing gives an explicit guardrail against learned over-filtering near the ground and sensor.

It is still a snow-removal method; rain, fog, road spray, de-icing mist or steam, dust, and multipath ghosts require separate validation.
Reflectivity and intensity distributions vary by LiDAR vendor, firmware, gain settings, target material, wetness, and range.
Rule-derived pseudo-labels can bake in the same sensor and weather assumptions that the learned model is meant to overcome.
Range-image projection can blur nonuniform vertical beam layouts and can lose detail where multiple 3D points collide into one pixel.
Close dynamic-object points, exhaust plumes, aircraft de-icing spray, reflective paint, and wet ground splash can look snow-like to range/intensity rules.
Post-processing improves stability but can hide systematic errors if it is tuned on a narrow sensor height or weather regime.
The paper is recent research, not a production safety argument.

Strong research fit for gates, service roads, aprons, and perimeter roads that must operate in snowfall.
Useful as a LiDAR preprocessing candidate before obstacle detection, mapping, and freespace extraction, especially when false snow obstacles cause nuisance stops.
Airside validation must include de-icing mist, glycol spray, jet blast disturbed snow, rain-on-snow, wet concrete, reflective aircraft skins, cones, chocks, personnel, and GSE.
It should not be the only adverse-weather mitigation; production stacks need radar, camera confidence handling, weather state estimation, health monitoring, and fallback behavior.
Best used with Radar-LiDAR Fusion in Adverse Weather so radar can challenge LiDAR-only filtering decisions.
Treat it as a candidate module feeding a larger Production Perception Systems design, not as standalone airside perception.

Preserve original point indices so downstream boxes, tracks, and map updates can be audited against removed points.
Calibrate intensity and reflectivity normalization per sensor SKU, firmware version, and operating mode.
Keep pseudo-label generation, post-processing thresholds, and model weights versioned together.
Log the raw point cloud, projected range image, predicted mask, post-processed mask, and removed point set for every regression scenario.
Add explicit test buckets for snow, rain, fog, road spray, de-icing mist, steam, dust, ghost/multipath returns, and dynamic-object points even if the model is trained only for snow.
Monitor over-removal around small obstacles, cone tips, personnel legs, tow bars, wing edges, and low-profile chocks.
Benchmark on embedded hardware; U-Net++ can meet nominal frame rates in the paper, but full system latency includes projection, post-processing, copying, and downstream perception.