TransFusion for Autoware (lidar_transfusion)

3D object detection model for LiDAR point clouds, used by the autoware_lidar_transfusion node in Autoware.

The model follows the TransFusion [1] architecture and, in this deployment, operates on LiDAR data only (x, y, z, intensity). It is exported as ONNX so it can be deployed across hardware; Autoware builds the TensorRT engine from the ONNX file on first launch.

Model overview

Model parameters from transfusion_ml_package.param.yaml:

Pre-processing (point cloud densification, voxelization) and post-processing (circle NMS, IoU-based NMS, yaw normalization, score thresholding) run in the node, not in the ONNX graph.

Files

TensorRT engines are not distributed here. TensorRT engines are specific to the GPU architecture and TensorRT version they are built on and are not portable, so Autoware builds them locally from the ONNX file on first launch (or via build_only:=true).

Inputs and outputs (as used by the node)

Input: ~/input/pointcloud (sensor_msgs/msg/PointCloud2). The node operates on raw cloud data and requires at least the following fields (additional fields are allowed):

[
  sensor_msgs.msg.PointField(name='x', offset=0, datatype=7, count=1),
  sensor_msgs.msg.PointField(name='y', offset=4, datatype=7, count=1),
  sensor_msgs.msg.PointField(name='z', offset=8, datatype=7, count=1),
  sensor_msgs.msg.PointField(name='intensity', offset=12, datatype=2, count=1)
]

Output: ~/output/objects (autoware_perception_msgs/msg/DetectedObjects): oriented 3D boxes with class and score. The node also publishes debug topics for cyclic time, pipeline latency, and per-stage processing times.

Usage in Autoware

The node expects these artifacts in ~/autoware_data/ml_models/lidar_transfusion/ (the launch file's default model_path) and launches with:

ros2 launch autoware_lidar_transfusion lidar_transfusion.launch.xml

Add build_only:=true to build the TensorRT engine from the ONNX as a one-off pre-task, and log_level:=debug for verbose logging. See the package README for the full parameter reference.

Training

The model was trained with MMDetection3D. According to the consuming package README, the TransFusion model of this family was trained on TIER IV's internal database (approximately 11k LiDAR frames) for 50 epochs; that statement is documented for the t4xx1_90m/v2 release, and no separate training notes are published for v2.1. The training configuration is not publicly documented.

Related implementations:

• CUDA-TransFusion: https://github.com/wep21/CUDA-TransFusion
• OpenPCDet: https://github.com/open-mmlab/OpenPCDet

Provenance

Consumers should pin the v2.1 revision when downloading, not main.

Limitations

• Only the five classes above are detected. Other road users fall outside the label set.
• The input point cloud must provide at least the x, y, z (float32) and intensity (uint8) fields in the layout shown above.
• Detections are limited to the configured point cloud range (about 92 m around the sensor origin).
• Trained on TIER IV internal data; accuracy on a different LiDAR setup (mounting position, beam count, concatenated clouds) can drop without fine-tuning.

Citation

@article{bai2022transfusion,
  title   = {TransFusion: Robust LiDAR-Camera Fusion for 3D Object Detection with Transformers},
  author  = {Bai, Xuyang and Hu, Zeyu and Zhu, Xinge and Huang, Qingqiu and Chen, Yilun and Fu, Hongbo and Tai, Chiew-Lan},
  journal = {arXiv preprint arXiv:2203.11496},
  year    = {2022}
}

References

• [1] Bai et al., "TransFusion: Robust LiDAR-Camera Fusion for 3D Object Detection with Transformers", arXiv:2203.11496, 2022.
• [2] CUDA-TransFusion: https://github.com/wep21/CUDA-TransFusion
• [3] MMDetection3D: https://github.com/open-mmlab/mmdetection3d
• [4] OpenPCDet: https://github.com/open-mmlab/OpenPCDet