File size: 3,114 Bytes
4ee8626
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
930a5ec
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
---
license: other
tags:
  - heal
  - horizon
  - lane-detection
---

# 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