OpenExploer/centerpoint_pointpillar

CenterPoint (PointPillars)

CenterPoint (PointPillars variant) voxelizes lidar point clouds into pillars, learns pillar features via PillarFeatureNet and scatters them into a 2D pseudo-image via PointPillarScatter, extracts multi-scale features via SECONDNeck, and CenterPointHead regresses box center, size, orientation, and velocity in an anchor-free manner to output 3D bounding boxes. Training uses CBGS (Class-Balanced Grouping and Sampling) data augmentation.

---

Deployment Metrics

Model Parameters

Model	Model Input	Backbone	Neck	Model Output
CenterPoint	lidar point cloud (B,N,5)	PointPillarScatter	SECONDNeck	3D bounding boxes (B,N,cls+reg)

Accuracy Metrics

March	Metric	float	calibration	qat	hbm
J6M	NDS	0.5865	0.5703	0.5853	0.5846
	mAP	0.474	0.4487	0.4699	0.4693

Results measured with march = March.NASH_M (J6M) configuration.

HEAL version: heal 0.0.2 / hbdk4-compiler 4.11.11 / horizon_plugin_pytorch 3.3.10.

Performance Metrics

Performance benchmark: FPS is measured with single-core 8 threads; latency is single-core single-thread; memory is peak DDR usage.

March	latency (ms)	fps	Memory Usage
J6M	9.23	183.98	51.10
J6P	7.56	833.52	47.20
J6B	-	-	-

J6B performance is not available for this model.

---

Model Overview

Core Design

CenterPoint (PointPillars variant) voxelizes lidar point clouds into pillars, learns pillar features via PillarFeatureNet and scatters them into a 2D pseudo-image via PointPillarScatter, extracts multi-scale features via SECONDNeck, and CenterPointHead regresses box center, size, orientation, and velocity in an anchor-free manner to output 3D bounding boxes. Training uses CBGS (Class-Balanced Grouping and Sampling) data augmentation.

• Task type: 3D object detection (3D Object Detection, lidar).
• backbone: PillarFeatureNet (num_input_features=5, num_filters=(64), learns pillar features) + PointPillarScatter (num_input_features=64, use_horizon_pillar_scatter=True, scatters pillar features into 2D pseudo-image).
• neck: SECONDNeck (in_feature_channel=64, down_layer_nums=[3,5,5], down_layer_channels=[64,128,256], up_layer_channels=[128,128,128], multi-scale feature extraction).
• Detection head: CenterPointHead (anchor-free, common_heads=dict(reg=(2,2), height=(1,2), dim=(3,2), rot=(2,2), vel=(2,2)), with_velocity=True).
• Loss function: CenterPointLoss (GaussianFocalLoss cls + L1Loss reg).
• Model input: lidar point cloud (B,N,5) (point_cloud_range=[-51.2,-51.2,-5.0,51.2,51.2,3.0], voxel_size=[0.2,0.2,8], max_num_points=20, max_voxels=(30000,40000), load_dim=5, use_dim=[0,1,2,3,4], num_sweeps=9).
• Model output: 10-class 3D bounding boxes (class_names = [car, truck, construction_vehicle, bus, trailer, barrier, motorcycle, bicycle, pedestrian, traffic_cone]), (B,N,cls+reg).

Official Repo and Paper

Official repo: https://github.com/tianweiy/CenterPoint Paper: https://arxiv.org/abs/2006.11275

Reference

For more J6 chip deployment details, see https://developer.horizon.auto/blog/10360