SLiDE

What It Is

SLiDE is an ECCV 2022 self-supervised LiDAR de-snowing method.
The name refers to self-supervised LiDAR de-snowing through reconstruction difficulty.
It removes snow points without requiring point-wise snow labels by learning which points are difficult to reconstruct from neighbors.
It is focused on snowfall, not a general adverse-weather cleanup method for rain, fog, road spray, dust, steam, or multipath.
It is closely related to LiSnowNet and later blind-spot variants such as 3D-KNN Blind-Spot Desnowing.

Snow points tend to have low spatial correlation with neighboring scene points.
Train a Point Reconstruction Network, PR-Net, to reconstruct a target point from its local neighborhood.
Train a Reconstruction Difficulty Network, RD-Net, to predict how difficult that reconstruction will be.
Treat high reconstruction difficulty as evidence that the point is likely snow/noise.
Use simple post-processing and a threshold on RD-Net output to classify snow points.
Extend the reconstruction task as a pretext task for label-efficient supervised de-snowing when limited labels are available.

Input: LiDAR point clouds represented so local neighborhoods can be sampled around target points.
Training input: unlabeled snowy point clouds; optional small labeled subsets for the semi-supervised extension.
Intermediate output: reconstructed target-point estimates from PR-Net.
Intermediate output: per-point reconstruction difficulty from RD-Net.
Output: point-wise snow/noise classifications after thresholding.
Non-output: object boxes, semantic classes beyond snow/noise, motion labels, or restored hidden geometry.

Sample target points and their neighborhoods from the point cloud.
Train PR-Net to infer each target point from neighboring points while withholding the target itself.
Use multi-hypothesis reconstruction so ambiguous local neighborhoods can have multiple plausible reconstructions.
Train RD-Net to predict the difficulty of PR-Net's reconstruction rather than directly using hand labels.
At test time, classify points with RD-Net output above a selected threshold as noise.
Apply post-processing to convert the difficulty estimate into a de-snowed point cloud.
In the semi-supervised variant, use reconstruction difficulty as a pretext signal to improve training efficiency when a small label set exists.

The arXiv page and ECCV paper present SLiDE as a label-free method comparable to fully supervised de-snowing.
The paper compares against ROR, DROR, and WeatherNet on synthesized snow-noise data.
Reported metrics include IoU, precision, and recall for binary snow/noise classification.
The ECCV paper reports a threshold of 2.9 for RD-Net output in its experiments.
It reports stronger label-free performance than DROR and comparable behavior to supervised WeatherNet in the studied setup.
The paper also shows that the semi-supervised extension improves label efficiency when only a small percentage of labels is available.

Does not require manual snow labels for the core self-supervised training path.
The reconstruction-difficulty signal is interpretable: snow is hard to reconstruct from coherent scene geometry.
Avoids relying solely on intensity, which varies across sensors, range, material, and weather exposure.
Can improve supervised training when labels are scarce.
Naturally highlights isolated or weakly correlated noise points.
Provides a strong conceptual bridge between classical neighborhood filters and learned point-cloud denoising.

Clean points that are isolated, thin, high, or geometrically unusual can also be hard to reconstruct and may be falsely removed.
Dynamic-object points, personnel limbs, tow bars, cones, cables, and aircraft edges can look like low-correlation outliers.
The method is snow-centric; rain streaks, dense fog attenuation, road spray, de-icing mist, dust, steam, and multipath ghosts have different statistics.
Performance depends on neighborhood construction and target-point sampling.
The threshold can be dataset-dependent and may drift across LiDAR models, beam layouts, and mounting heights.
It does not reason over time, so it can confuse transient weather with transient object motion.
It remains a research method and needs production wrappers, monitoring, and fallback logic.

Useful for researching snowfall cleanup where explicit snow labels are expensive or inconsistent.
Riskier near aircraft stands because small but safety-critical objects can be locally sparse and hard to reconstruct.
Needs airport-specific false-positive analysis on cones, chocks, belt-loader edges, tow bars, baggage carts, personnel, and aircraft extremities.
Does not directly handle de-icing spray, jet blast snow clouds, rain spray, or steam unless retrained and evaluated for those artifacts.
Should be integrated with Radar-LiDAR Fusion in Adverse Weather for cross-sensor checks.
Use the deployment discipline in Production Perception Systems before putting it in any safety-relevant path.

Keep PR-Net and RD-Net artifacts versioned together; RD-Net's target depends on PR-Net behavior.
Store the reconstruction difficulty map for replay debugging, not just the final removed points.
Treat the threshold as a calibration parameter that must be validated by route, sensor, and weather severity.
Add scenario tests where clean sparse points are intentionally present so over-filtering is visible.
Compare against LiSnowNet, LIORNet, and 3D-OutDet on the same point-index-preserving evaluation harness.
Evaluate downstream detection and tracking impact, not just snow IoU.
Keep raw point clouds available for safety analysis because the method can delete evidence of small obstacles.