AdverseNet

What It Is

AdverseNet is an IEEE Sensors Journal 2025 LiDAR point-cloud denoising network for rainy, snowy, and foggy weather.
It is a unified learned denoiser rather than a snow-only filter.
The official repository describes it as using Cylindrical Tri-Perspective View, CTPV, representation and a two-stage training strategy.
It targets point-wise removal of weather noise points from LiDAR clouds before downstream perception.
It is relevant to broader adverse-weather stacks alongside TripleMixer, 3D-OutDet, and DenoiseCP-Net.

Represent the point cloud with CTPV so the network can process cylindrical and tri-perspective structure.
Learn generic adverse-weather features first across rain, snow, and fog.
Fine-tune or specialize in a second stage for weather-specific features.
Use a unified model rather than separate filters for each weather type.
Train/evaluate on DENSE and SnowyKITTI to cover controlled rain/fog and snow-corrupted point clouds.
Output denoising labels that suppress weather artifacts while retaining scene points.

Input: LiDAR point clouds with coordinate and intensity-like features, converted into AdverseNet's CTPV representation.
Training input: DENSE and SnowyKITTI data after repository-specific format conversion and invalid-point filtering.
Intermediate output: stage-1 model weights that learn generic weather-noise features.
Intermediate output: stage-2 model weights specialized by weather-feature learning.
Output: point-wise denoising segmentation for rain, snow, and fog artifacts.
Non-output: no object detector, no cooperative perception message, no radar evidence, and no explicit restoration of missing returns.

Convert dataset formats into the repository's text representation and filter invalid points.
Construct the CTPV representation from the point cloud.
Train stage 1 to learn shared features across adverse-weather noise types.
Train stage 2 with weather-specific objectives or flags to improve specialization.
Run segmentation-style evaluation to classify and remove weather noise points.
Feed the denoised cloud to downstream modules only after preserving the removed-point mask for audit.

The official repository links the IEEE Sensors Journal paper and includes training scripts for stage 1 and stage 2.
The paper citation is IEEE Sensors Journal, volume 25, number 5, pages 8950-8961, 2025.
The repository reports comparative experiments on DENSE and SnowyKITTI.
The reported mean IoU values are 94.67 percent on DENSE and 99.33 percent on SnowyKITTI.
The setup depends on PyTorch, mmdetection3d, mmcv, mmdet, mmsegmentation, torch_scatter, spconv, and related packages.
Results are best interpreted as point-wise denoising benchmark results, not full autonomous-driving safety validation.

Explicitly covers three major LiDAR weather degradations: rain, snow, and fog.
Two-stage training separates generic weather-noise structure from weather-specific cues.
Public code, configs, weights, and scripts make replication more practical than paper-only methods.
CTPV is designed to preserve more geometric context than a single flat projection.
DENSE plus SnowyKITTI gives broader weather coverage than snow-only WADS evaluation.
It can act as a learned preprocessing module before detection, segmentation, mapping, or tracking.

DENSE and SnowyKITTI do not cover every real airport artifact, especially de-icing mist, glycol spray, steam, dust, jet blast snow, and wet apron splash.
Fog often causes attenuation and missing returns, not just removable foreground points.
Controlled-weather or synthetic datasets can overstate transfer to real mixed-weather operations.
The model may remove sparse valid objects if they resemble weather noise in CTPV space.
Dynamic-object points and multipath ghosts are not the same class as weather particles.
The dependency stack is heavy and may be difficult to integrate into low-latency embedded perception without simplification.
It is not production-ready without calibration, runtime monitoring, domain validation, and fallback behavior.

Strong candidate for research when rain, snow, and fog must be handled by one LiDAR denoising module.
Needs additional airside data for de-icing operations, aircraft reflections, service-road spray, night lighting, and wet concrete.
CTPV may help around complex 3D geometry, but airport-specific thin and low objects require targeted false-positive tests.
Should be combined with Radar-LiDAR Fusion in Adverse Weather rather than trusted as a LiDAR-only decision maker.
Use Production Perception Systems as the deployment gate for monitoring, redundancy, and safety fallback.
Compare with TripleMixer for benchmark breadth and with 3D-OutDet for lower-memory filtering.

Reproduce the repository's data conversion and invalid-point filtering exactly before comparing metrics.
Track stage-1 and stage-2 weights, weather flags, and lambda values as part of model versioning.
Preserve raw and denoised clouds plus removed masks for every replay.
Benchmark the full mmdetection3d/spconv pipeline on target hardware before planning real-time use.
Add airport-specific evaluation classes for snow, rain, fog, road spray, de-icing mist/steam, dust, multipath ghosts, and dynamic objects.
Monitor downstream performance separately; high denoising mIoU can still degrade detection of rare low-profile hazards.
Avoid using DENSE/SnowyKITTI metrics as direct evidence for apron production readiness.