lidar_centerpoint is a package for detecting dynamic 3D objects.
In this implementation, CenterPoint [1] uses a PointPillars-based [2] network to inference with TensorRT.
We trained the models using https://github.com/open-mmlab/mmdetection3d.
| Name | Type | Description |
|---|---|---|
~/input/pointcloud |
sensor_msgs::msg::PointCloud2 |
input pointcloud |
| Name | Type | Description |
|---|---|---|
~/output/objects |
autoware_auto_perception_msgs::msg::DetectedObjects |
detected objects |
debug/cyclic_time_ms |
tier4_debug_msgs::msg::Float64Stamped |
cyclic time (msg) |
debug/processing_time_ms |
tier4_debug_msgs::msg::Float64Stamped |
processing time (ms) |
| Name | Type | Default Value | Description |
|---|---|---|---|
score_threshold |
float | 0.4 |
detected objects with score less than threshold are ignored |
densification_world_frame_id |
string | map |
the world frame id to fuse multi-frame pointcloud |
densification_num_past_frames |
int | 1 |
the number of past frames to fuse with the current frame |
trt_precision |
string | fp16 |
TensorRT inference precision: fp32 or fp16 |
encoder_onnx_path |
string | "" |
path to VoxelFeatureEncoder ONNX file |
encoder_engine_path |
string | "" |
path to VoxelFeatureEncoder TensorRT Engine file |
head_onnx_path |
string | "" |
path to DetectionHead ONNX file |
head_engine_path |
string | "" |
path to DetectionHead TensorRT Engine file |
nms_iou_target_class_names |
list[string] | - | target classes for IoU-based Non Maximum Suppression |
nms_iou_search_distance_2d |
double | - | If two objects are farther than the value, NMS isn't applied. |
nms_iou_threshold |
double | - | IoU threshold for the IoU-based Non Maximum Suppression |
- The
object.existence_probabilityis stored the value of classification confidence of a DNN, not probability.
You can download the onnx format of trained models by clicking on the links below.
| Name | URLs | Description |
|---|---|---|
centerpoint |
pts_voxel_encoder pts_backbone_neck_head |
There is a single change due to the limitation in the implementation of this package. num_filters=[32, 32] of PillarFeatureNet |
centerpoint_tiny |
pts_voxel_encoder pts_backbone_neck_head |
The same model as default of v0. |
These changes are compared with this configuration.
| Name | URLs | Description |
|---|---|---|
default |
pts_voxel_encoder pts_backbone_neck_head |
There are two changes from the original CenterPoint architecture. num_filters=[32] of PillarFeatureNet and ds_layer_strides=[2, 2, 2] of RPN |
[1] Yin, Tianwei, Xingyi Zhou, and Philipp Krähenbühl. "Center-based 3d object detection and tracking." arXiv preprint arXiv:2006.11275 (2020).
[2] Lang, Alex H., et al. "Pointpillars: Fast encoders for object detection from point clouds." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019.
[3] https://github.com/tianweiy/CenterPoint
[4] https://github.com/open-mmlab/mmdetection3d
[5] https://github.com/open-mmlab/OpenPCDet