LiSnowNet

What It Is

LiSnowNet is an IROS 2022 unsupervised LiDAR snow-removal method for point clouds corrupted by snowfall.
It was designed to avoid slow nearest-neighbor snow filters such as DROR and DSOR on modern high-point-count LiDAR sweeps.
It converts point clouds into range images and trains a CNN to output a residual/noise image without point-wise snow labels.
It is snow-specific in its formulation and should not be treated as a universal rain, fog, spray, dust, or multipath-removal network.
It is a useful baseline for newer self-supervised approaches such as SLiDE, 3D-KNN Blind-Spot Desnowing, and LIORNet.

Project a LiDAR sweep into a two-channel range image: distance and intensity.
Assume clean range images are sparse under frequency and wavelet transforms, while snow corruption makes the representation less sparse.
Train a CNN to predict the residual image representing the weather/noise component.
Use sparsity losses based on FFT and DWT coefficients instead of supervised snow labels.
Subtract the predicted residual from the input range image to estimate a cleaner scene representation.
Classify snow using residual-space conditions that encode whether a point is foreground, low intensity, and not sparse.

Input: a single LiDAR point cloud with per-point coordinates and intensity.
Preprocessed input: a spherical range image with distance and intensity channels.
Training input: unlabeled point clouds from CADC and WADS.
Output: a residual image estimating the snow/noise component.
Output: a denoised point cloud containing points predicted as non-snow.
Non-output: no semantic object classes, no rain/fog class separation, no temporal tracks, and no explicit uncertainty map.

Convert each point to range, inclination, and azimuth, then discretize into a panoramic range image.
Scale range and intensity so close snow-like returns matter despite the larger maximum LiDAR range.
Fill void pixels with a limited preprocessing chain; valid measurement pixels are not modified by that void-filling step.
Use a modified MWCNN-style architecture with residual blocks, circular convolutions, dropout, and reduced channel counts for real-time operation.
Replace pooling and transposed convolution with DWT downsampling and inverse DWT upsampling.
Train the network to produce the residual/noise image rather than directly producing a hard binary mask.
Apply a residual-space decision boundary after training to decide which original points are snow.

The paper trains on CADC and WADS without point-wise labels, using sequence-level train/validation splits.
WADS labels are used for evaluation because it provides point-wise snow labels.
Reported baselines include DROR, DSOR, and a median filter.
The paper reports LiSnowNet variants with higher IoU than the nearest-neighbor filters and roughly 6.8 ms runtime on the tested desktop GPU.
The arXiv abstract reports about 52x speedup versus prior state-of-the-art nearest-neighbor methods.
The paper also shows cleaner CADC maps when denoised point clouds are fed into mapping with RTK poses or LeGO-LOAM.

Does not require point-wise snow labels for training.
Avoids per-point nearest-neighbor search at inference by using image-like operations on a range projection.
FFT/DWT sparsity gives a clear signal-processing rationale for the loss.
The runtime profile is attractive for 10 Hz LiDAR and GPU deployment.
Works naturally as a preprocessing block before mapping or object detection.
Public code and pretrained artifacts make it a practical research baseline.

It is tuned around snowflake behavior in range and intensity images; rain, fog, road spray, de-icing mist, dust, and steam are not covered by the main evidence.
Snow-like small foreground structures can be over-removed when they appear sparse, close, or low intensity.
Range-image projection can lose 3D adjacency, especially near depth discontinuities, thin structures, and multi-return ambiguity.
Static sparsity assumptions may fail around dense vegetation, chain-link fences, aircraft edges, jet bridges, and reflective wet surfaces.
Intensity behavior changes across LiDAR vendors and weather exposure.
It has no built-in temporal reasoning to distinguish moving objects from transient snow.
It is research code, not a production-certified filter.

Good candidate baseline for snow removal before mapping on service roads, aprons, and perimeter routes.
Less appropriate as a sole filter around aircraft stands where de-icing mist, exhaust, wet concrete spray, and reflective metal are common.
Needs airport-specific data before use near chocks, cones, tow bars, personnel, belt loaders, dollies, and low-profile obstacles.
Pair with radar-based checks and weather state estimation as described in Radar-LiDAR Fusion in Adverse Weather.
For deployment decisions, follow the redundancy and monitoring guidance in Production Perception Systems.
Compared with LIORNet, LiSnowNet is simpler and faster but has fewer explicit physics/rule safeguards after inference.

Keep the spherical projection parameters tied to the exact LiDAR vertical beam layout and horizontal resolution.
Preserve a reversible mapping from range-image pixels to original point indices.
Validate the void-pixel filling path separately; it should not silently modify valid returns.
Refit or verify residual thresholds when changing LiDAR model, intensity scaling, or mounting height.
Track removed points by semantic bucket in evaluation: snow, rain, fog, spray, mist/steam, dust, multipath ghosts, and dynamic-object points.
Benchmark end-to-end latency including projection and reprojection, not only CNN inference.
Use it as a baseline against SLiDE, 3D-OutDet, TripleMixer, and LIORNet.