README.md

metadata

license: other
tags:
  - heal
  - horizon
  - bev
  - lane-detection

MapTR+HENet (BevFormer)

MapTR uses HENet as the camera backbone to extract multi-view features, transforms them to BEV features via BevFormer's single-frame ViewTransformer and BEV Encoder, then feeds BEV features to the MapTR decoder with fixed-point polyline queries (fixed_ptsnum_per_pred_line=20) to predict vectorized map elements (divider/ped_crossing/boundary). This task has use_lidar_gt=False; map GT is generated online.

---

Deployment Metrics

Model Parameters

Model	Model Input	Backbone	Neck	Model Output
MapTR	6-camera multi-view images (B,6,3,480,800)	HENet-tiny	FPN	vectorized map (B,L,P,2)

Accuracy Metrics

March	Metric	float	calibration	qat	hbm
J6M	chamfer mAP (MAP)	0.6626	0.6588	—	0.6315

Data tested with march = March.NASH_M (J6M); this task has no QAT stage (— in the qat column).

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

Performance Metrics

Performance test methodology: FPS is measured with 8 threads on a single core; Latency is measured with single core, single thread; Memory is peak DDR usage.

March	latency (ms)	fps	Memory Usage
J6M	9.30	111.22	88.40
J6P	6.16	663.68	83.30
J6B	33.28	30.85	80.00

---

Model Overview

Core Design

MapTR uses HENet as the camera backbone to extract multi-view features, transforms them to BEV features via BevFormer's single-frame ViewTransformer and BEV Encoder, then feeds BEV features to the MapTR decoder with fixed-point polyline queries (fixed_ptsnum_per_pred_line=20) to predict vectorized map elements (divider/ped_crossing/boundary). This task has use_lidar_gt=False; map GT is generated online.

• Task type: Online Vectorized Map Construction.
• backbone: HENet-tiny (pretrained), extracts multi-view camera features.
• neck: FPN.
• view transformation: SingleFrameBevFormerViewTransformer + SingleFrameBEVFormerEncoder (queue_length=1/test_queue_length=1).
• map elements: map_classes=[divider, ped_crossing, boundary], fixed_ptsnum_per_gt_line=20.
• BEV range: use_lidar_gt=False else branch, point_cloud_range=[-30.0,-15.0,-10.0,30.0,15.0,10.0], bev_h_=50, bev_w_=100 (bev 50×100).
• Model input: 6-camera single-frame images (B,6,3,480,800) .
• Model output: 3 classes of vectorized map elements (divider/ped_crossing/boundary), 20 points per line.

Official Repo and Paper

Official repo: https://github.com/hustvl/MapTR Paper: https://arxiv.org/abs/2208.14437

Note: The camera backbone HENet is developed in HEAL; the official repo uses a different backbone.

Reference

For more J6 chip deployment details, see https://developer.horizon.auto/blog/14100