OpenExplorer/maptr_sparse_henet_tinym

SparseMapTR+HENet

SparseMapTR uses HENet as the camera backbone to extract multi-view features, converts them to BEV features, then feeds them to SparseMapHead (6-layer sparse query stack) for vectorized map element prediction. Unlike MapTR, SparseMapTR uses a sparse query mechanism (InstanceBankOE + SparsePoint3DEncoder), refining only a small set of candidate queries iteratively to reduce compute. This task has use_lidar_gt=True.

---

Deployment Metrics

Model Parameters

Model	Model Input	Backbone	Neck	Model Output
SparseMapTR	6-camera multi-view images (B,6,3,256,704) + lidar point cloud (B,N,5)	HENet-tiny	FPN	vectorized map (B,L,P,2)

Accuracy Metrics

March	Metric	float	calibration	qat	hbm
J6M	chamfer mAP (MAP)	0.5924	0.5882	—	0.5892

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 for J6M/J6P is measured with 8 threads on a single core; J6B uses 4 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	11.46	89.46	68.40
J6P	9.32	259.76	98.80
J6B	160.03	17.03	120.00

---

Model Overview

Core Design

SparseMapTR uses HENet as the camera backbone to extract multi-view features, converts them to BEV features, then feeds them to SparseMapHead (6-layer sparse query stack) for vectorized map element prediction. Unlike MapTR, SparseMapTR uses a sparse query mechanism (InstanceBankOE + SparsePoint3DEncoder), refining only a small set of candidate queries iteratively to reduce compute. This task has use_lidar_gt=True.

• Task type: Sparse Vectorized Map Construction.
• backbone: HENet-tiny (pretrained), extracts multi-view camera features.
• neck: FPN (out_strides=[4,8,16,32], outputs 256-dim multi-scale features).
• decoder: SparseMapHead (6-layer sparse query stack, InstanceBankOE + SparsePoint3DEncoder).
• map elements: map_classes=[divider, ped_crossing, boundary], fixed_ptsnum_per_gt_line=20.
• BEV range: use_lidar_gt=True branch, point_cloud_range=[-15.0,-30.0,-10.0,15.0,30.0,10.0], bev_h_=100, bev_w_=50 (bev 100×50).
• Model input: 6-camera images (B,6,3,256,704) + lidar point cloud (B,N,D) (use_lidar_gt=True).
• 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