3D-OutDet

What It Is

3D-OutDet is a learned outlier detector for 3D LiDAR point clouds captured in adverse weather.
It targets precipitation outliers, especially falling snow and rain spray, with a focus on low memory footprint and fast execution.
The project page shows de-snowing and rain-spray-removal examples.
It operates directly on 3D point clouds rather than relying only on 2D range images.
It is a useful comparator for LiSnowNet, TripleMixer, and AdverseNet.

Use a lightweight neural architecture built around a neighborhood convolution that processes nearest neighbors only.
Limit the model depth and computation so adverse-weather filtering can run on onboard systems.
Precompute or compute KNN neighborhoods, then apply learned local feature aggregation to classify outliers.
Trade a small amount of accuracy for substantial reductions in memory, operations, and execution time compared with heavier point-cloud networks.
Treat noisy precipitation points as a binary outlier class that can be removed before downstream perception.
Preserve enough 3D geometry to avoid some range-image projection failures.

Input: raw 3D LiDAR point clouds and KNN neighborhood information.
Training input: labeled outlier datasets such as WADS, SnowyKITTI, and SemanticSpray.
Intermediate output: nearest-neighbor features processed by the 3D-OutDet modules.
Output: point-wise outlier predictions for snow or rain spray removal.
Output: denoised point clouds after removing predicted outliers.
Non-output: no object boxes, no weather classification beyond the trained outlier classes, no radar fusion, and no restored hidden surfaces.

Remove duplicated points where needed, especially in WADS preprocessing, because duplicates distort nearest-neighbor neighborhoods.
Generate or load KNN distances and neighbor indices for training and evaluation.
Train a binary outlier detector for the selected dataset/weather target.
Apply the learned neighborhood convolution on local point neighborhoods.
Produce point-wise outlier labels and filter predicted noise points.
Evaluate both direct denoising quality and downstream semantic segmentation effects.

The paper reports experiments on WADS, SnowyKITTI, and SemanticSpray.
The official repository includes separate WADS, KITTI desnowing, and spray-filtering training/evaluation scripts.
The project reports competitive WADS performance with much lower memory and computation than larger baselines.
The paper abstract reports a 0.16 percent mIoU sacrifice while reducing memory by 99.92 percent, operations by 96.87 percent, and execution time by 82.84 percent per point cloud on WADS.
The repository reports a WADS benchmark update with RandLA-Net and 3D-OutDet, including 3D-OutDet precision 96.78, recall 92.76, F1 94.73, and mIoU 90.00.
The authors note that a C++/TensorRT implementation is underway, which indicates deployment interest but not production readiness.

More memory-conscious than many point-cloud denoising networks.
Direct 3D processing avoids some failure modes of spherical range projections.
Covers both de-snowing and rain-spray-removal examples.
Public code includes dataset preprocessing, training, evaluation, and sample configs.
Explicit duplicate-removal warning for WADS is practically useful for fair benchmarking.
Can improve downstream semantic segmentation by removing weather outliers before the main model.

KNN preprocessing is a dependency and can be distorted by duplicates, uneven density, ego-motion, or different scan patterns.
It is trained as an outlier detector, so rare but valid sparse objects can be removed.
Fog attenuation, de-icing mist, steam, dust, and multipath ghosts may not match the snow/spray training labels.
Binary filtering does not distinguish weather noise from dynamic-object points unless the training data teaches that boundary.
Removing points before tracking can make a downstream system overconfident if filtered points are not auditable.
The public repository has research scripts and no release package, so integration maturity is limited.
Production claims should be treated cautiously until embedded implementation, monitoring, and fail-safe behavior are demonstrated.

Relevant for airport road spray, wet apron splash, and snowfall because those resemble the public examples more than pure fog.
Needs careful false-positive evaluation on low-profile airside obstacles: chocks, cones, tow bars, dollies, cables, and personnel.
SemanticSpray makes it more interesting for airside than snow-only methods, but de-icing mist and jet blast snow are still separate artifacts.
Use radar and camera corroboration as described in Radar-LiDAR Fusion in Adverse Weather.
Follow Production Perception Systems for audit logging, health checks, and fallback behavior.
Compare with TripleMixer for broader snow/fog/rain coverage and with AdverseNet for unified rain/snow/fog denoising.

Make duplicate-point removal deterministic and record it separately from learned denoising.
Cache KNN artifacts only when the exact point cloud, projection, and preprocessing version are fixed.
Preserve point indices and removed-point masks for replay and incident analysis.
Profile KNN search, model inference, and post-filtering separately on target hardware.
Maintain separate configs for snow, spray, and any future mist/dust use case.
Add regression tests for sparse valid objects and moving actors so the outlier detector does not become a small-object remover.
Treat the C++/TensorRT path as future integration work unless it is available and verified in the target stack.