OpenExplorer/bev_vtv2_resnet50

BEV (ResNet-50)

This model follows the LSS (Lift-Splat-Shoot) view transformation approach: 6 camera images are processed by ResNet-50 + FPN for multi-scale features, projected to the BEV plane via LSSTransformer, encoded by BevEncoder, and finally predicted on the BEV grid by CenterPointHead for 3D bounding boxes.

---

Deployment Metrics

Model Parameters

Model	Model Input	Backbone	Neck	Model Output
BEV-VTv2	6-camera multi-view images (B,6,3,256,704)	ResNet-50	FPN	BEV grid 3D bounding boxes (B,N,cls+reg)

Accuracy Metrics

March	Metric	float	calibration	qat	hbm
J6M	NDS	0.3603	0.3546	0.3564	0.3539
	mAP	0.2833	0.2799	0.2805	0.2797

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	16.45	62.04	121.20
J6P	11.42	347.23	123.10
J6B	-	-	-

J6B performance is not available for this model.

---

Model Overview

Core Design

This model follows the LSS (Lift-Splat-Shoot) view transformation approach: 6 camera images are processed by ResNet-50 + FPN for multi-scale features, projected to the BEV plane via LSSTransformer, encoded by BevEncoder, and finally predicted on the BEV grid by CenterPointHead for 3D bounding boxes.

• Task type: BEV 3D object detection (BEV 3D Object Detection).
• backbone: ResNet-50 (ResNet50, include_top=False, pretrained num_classes=1000).
• neck: FPN (FPN, multi-scale feature pyramid, output strides 16/32, fix_out_channel=256).
• Detection head: CenterPointHead (BEV grid center-point detection head).
• Loss function: CenterPointLoss (GaussianFocalLoss + L1Loss).
• Model input: 6-camera multi-view images, (B,6,3,256,704) (original orig_shape=(3,900,1600) → resize (3,396,704) → crop data_shape=(3,256,704), num_views=6).
• Model output: 3D bounding boxes on BEV grid (class + center + size + orientation), CenterPointHead outputs heatmap + reg/height/dim/rot/vel per task group, num_classes=10, decoded via CenterPointPostProcess + NMS.

Official Repo and Paper

Official repo: https://github.com/HuangJunJie2017/BEVDet Paper: https://arxiv.org/pdf/2112.11790

Note: Backbone is ResNet-50; detection head is CenterPoint.

Reference

For more J6 chip deployment details, see https://developer.horizon.auto/blog/14085