Skip to content

FRNet

FRNet is a LiDAR-based 3D semantic segmentation model built on frustum and range-view feature fusion. It is integrated under the segmentation3d task namespace for NuScenes and T4Dataset.

Summary

Property Value
Task 3D semantic segmentation
Modality LiDAR
Input Point cloud
Output Point-wise semantic class labels
Architecture Frustum encoder, FRNet backbone, segmentation heads
Datasets NuScenes, T4Dataset

Available Configurations

Config Name Dataset Purpose
segmentation3d/frnet/nuscenes/hdl32e_b16_150k_nuscenes NuScenes Standard NuScenes configuration
segmentation3d/frnet/t4dataset/base_ot128 T4Dataset T4Dataset OT128 configuration
segmentation3d/frnet/t4dataset/base_qt128 T4Dataset T4Dataset QT128 configuration

The T4Dataset rows are bases. They declare the classes the model trains and require the taxonomy that folds the labels of the database onto them, so a T4Dataset experiment that binds one lives in an external config root, see Configuration.

Training

autoware-ml train --config-name segmentation3d/frnet/nuscenes/hdl32e_b16_150k_nuscenes

For a pipeline validation run:

autoware-ml train \
    --config-name segmentation3d/frnet/nuscenes/hdl32e_b16_150k_nuscenes \
    +trainer.fast_dev_run=true

Evaluation

autoware-ml test \
    --config-name segmentation3d/frnet/nuscenes/hdl32e_b16_150k_nuscenes \
    --weights mlruns/frnet/NuScenesDataset-base-v1.0.0/nuscenes/hdl32e_b16_150k_nuscenes/<run_id>/artifacts/checkpoints/best.ckpt

Deployment

autoware-ml deploy \
    --config-name segmentation3d/frnet/nuscenes/hdl32e_b16_150k_nuscenes \
    --weights mlruns/frnet/NuScenesDataset-base-v1.0.0/nuscenes/hdl32e_b16_150k_nuscenes/<run_id>/artifacts/checkpoints/best.ckpt

To validate ONNX export without building a TensorRT engine:

autoware-ml deploy \
    --config-name segmentation3d/frnet/nuscenes/hdl32e_b16_150k_nuscenes \
    --weights mlruns/frnet/NuScenesDataset-base-v1.0.0/nuscenes/hdl32e_b16_150k_nuscenes/<run_id>/artifacts/checkpoints/best.ckpt \
    deploy.tensorrt.enabled=false

The exported ONNX model returns point-wise semantic probabilities through a final softmax layer. Training and evaluation continue to use logits inside the Lightning model.

Deployment uses an explicit FRNet export wrapper with copied model submodules and explicit single-sample export metadata, so export-specific behavior does not mutate the training model.

Data Pipeline

The FRNet preprocessing path converts raw point clouds into frustum and range-view representations and prepares both point-level and range-view supervision targets. The training pipeline includes dataset-specific augmentations and FRNet-specific transforms such as frustum mixing, instance copy, and range interpolation.

The standard FRNet training configs follow the AWML experiment contract:

  • AdamW with OneCycleLR
  • step-based validation every 1500 training steps
  • best-checkpoint selection by validation loss
  • mix augmentations that apply a secondary-sample transform pipeline before frustum mixing and instance copy

Implementation

Path Description
autoware_ml/models/segmentation3d/frnet.py FRNet Lightning model wrapper
autoware_ml/models/segmentation3d/encoders/frnet.py Frustum feature encoder
autoware_ml/models/segmentation3d/backbones/frnet.py FRNet backbone
autoware_ml/models/segmentation3d/heads/frnet.py FRNet segmentation heads
autoware_ml/losses/segmentation3d/ Segmentation losses used by FRNet
autoware_ml/datamodule/t4dataset/dataset.py T4Dataset records as a dataset
autoware_ml/datamodule/nuscenes/dataset.py nuScenes dataset
autoware_ml/datamodule/t4dataset/segmentation3d.py Semantic masks of a T4Dataset record
autoware_ml/transforms/point_cloud/ Point cloud loading and augmentation used by FRNet
autoware_ml/preprocessing/segmentation3d/frustum_range.py Frustum and range preprocessing
autoware_ml/configs/tasks/segmentation3d/frnet/ Experiment configurations

Acknowledgment

The Autoware-ML FRNet implementation was ported from the official FRNet project.

  • Repository: https://github.com/Xiangxu-0103/FRNet
  • License: Apache 2.0
  • Paper: Xu, Xiang, et al. "FRNet: Frustum-Range Networks for Scalable LiDAR Segmentation." IEEE Transactions on Image Processing, vol. 34, pp. 2173-2186, 2025.