OpenExplorer/petr_efficientnetb3

PETR (EfficientNet-b3)

PETR introduces 3D position into Transformer decoding: EfficientNet-b3 extracts multi-view image features, which are associated with 3D spatial positions via 3D positional encoding (SinePositionalEncoding3D); PETRTransformer queries interact directly with 3D position-encoded features to predict 3D detection boxes, without explicit BEV feature construction.

---

Deployment Metrics

Model Parameters

Model	Model Input	Backbone	Neck	Model Output
PETR	6-camera multi-view images (B,6,3,512,1408)	EfficientNet-b3	—	Detection boxes (B,900,21)

Accuracy Metrics

March	Metric	float	calibration	qat	hbm
J6M	NDS	0.3881	0.3679	0.38	0.38
	mAP	0.3031	0.2807	0.2942	0.2942

Data 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 test methodology: FPS is measured with single-core eight-thread; Latency is measured with single-core single-thread; Memory is peak DDR usage.

March	latency (ms)	fps	Memory Usage
J6M	33.13	30.51	101.00
J6P	21.56	186.12	105.30
J6B	-	-	-

J6B performance is not available for this model.

---

Model Overview

Core Design

PETR introduces 3D position into Transformer decoding: EfficientNet-b3 extracts multi-view image features, which are associated with 3D spatial positions via 3D positional encoding (SinePositionalEncoding3D); PETRTransformer queries interact directly with 3D position-encoded features to predict 3D detection boxes, without explicit BEV feature construction.

• Task type: BEV 3D object detection (BEV 3D Object Detection).
• backbone: EfficientNet-b3 (efficientnet, model_type=b3, include_top=False removes the classification head).
• neck: — (no standalone neck; backbone features feed directly into PETRHead).
• Detection head: PETRHead + PETRTransformer + PETRDecoder (3D position-aware Transformer decoder, num_query=900, num_views=6, embed_dims=256).
• Loss function: FocalLoss (cls) + L1Loss (reg), with Detr3dTarget Hungarian matching.
• Model input: 6-camera multi-view images, data_shape=(3,512,1408) (original (3,900,1600) resized to (3,792,1408) then cropped to data_shape), 6 views stacked as (B,6,3,512,1408).
• Model output: Detection boxes (B,900,21) — 900 queries × (10 class scores + 11 regression: center xyz + size wlh + orientation sin/cos + velocity vxvy), num_classes=10, reg_out_channels=10.

Official Repo and Paper

Official repo: https://github.com/megvii-research/PETR Paper: https://arxiv.org/abs/2203.05625

Note: The backbone is EfficientNet-b3; the official repo uses a different backbone.

Reference

For more J6 chip deployment details, see https://developer.horizon.auto/blog/14091