jiacheng.lin
1bde0fe430
- Implemented class name loading for ADAS campaigns in `_class_names_for_campaign`. - Enhanced annotation parsing to support cuboid and 2D rectangle labels in `_parse_ls_annotations`. - Updated catalog building to include ADAS packs and batches in `build_catalog_signature`. - Added ADAS cuboid promotion logic in `adas_cuboid.py`, supporting both quaternion_json and YOLO HBB formats. - Introduced validation for ADAS batches to check for YOLO HBB labels in `adas_cuboid.py`. - Modified delivery scanning to include ADAS project deliveries in `scan.py`. - Extended job execution to handle ADAS exports for YOLO format in `runner.py`. - Updated labeling export logic to accommodate ADAS project in `export_cuboid_batch.py`. - Enhanced format conversion to support rectangle labels in `format_converter.py`. - Improved CVAT task handling in `service.py` to ensure task ID updates on each labeling job. - Updated web API endpoints to reflect ADAS project changes in `hsap-api.ts`. - Added ADAS catalog types and UI components for displaying ADAS packs and batches in `dmsCatalog.ts`, `CatalogPage.tsx`, and `ExportPage.tsx`. - Adjusted workflow registry to align with new ADAS project structure.
79 lines
2.5 KiB
Markdown
79 lines
2.5 KiB
Markdown
# ADAS 前向训练包 `adas_v1`
|
||
|
||
> **2026-07 更新**：pack 已从 `adas_moon3d_v1` 重命名为 `adas_v1`（见 `workflow.registry.yaml`）。
|
||
> 新增 YOLO HBB 导出路径和 2D 矩形标注支持。
|
||
|
||
## 目录结构
|
||
|
||
```text
|
||
datasets/adas/packs/adas_v1/
|
||
├── sources/{batch}/
|
||
│ ├── images/
|
||
│ ├── calib/
|
||
│ └── labels/
|
||
│ ├── quaternion_json/ # 3D cuboid 导出（有 calib 时优先）
|
||
│ └── yolo-hbb/ # 2D YOLO HBB 导出（无 calib 回退）
|
||
├── lists/
|
||
│ ├── train_stems.txt
|
||
│ └── val_stems.txt
|
||
└── manifests/
|
||
└── pack_index.yaml
|
||
```
|
||
|
||
## class_id（BK2/MOON 顺序）
|
||
|
||
| ID | 类别 |
|
||
|----|------|
|
||
| 0 | pedestrian |
|
||
| 1 | car |
|
||
| 2 | truck |
|
||
| 3 | bus |
|
||
| 4 | motorcycle |
|
||
| 5 | tricycle |
|
||
| 6 | traffic cone |
|
||
|
||
定义于 [`data/送标/adas/adas.registry.yaml`](../../data/送标/adas/adas.registry.yaml) 与 [`datasets/labeling.registry.yaml`](../datasets/labeling.registry.yaml)。
|
||
|
||
## 管线
|
||
|
||
1. **labeling_export** — CVAT ls_annotations → `labels/quaternion_json/*.json`
|
||
- 支持 **cuboid**（3D 框）和 **rectanglelabels**（2D 矩形）两种标注类型
|
||
- 2D 矩形自动转为 quaternion_json detection（`fit_ok=false`, `source="2d_rect"`）
|
||
2. **cuboid_fit_3d**（有 calib 时自动触发）— 补全 3D 字段
|
||
3. **build_adas**（审核）— `promote_batch` 复制到 pack + 刷新 stem 列表
|
||
- 优先使用 quaternion_json，无 cuboid 标注时回退到 YOLO HBB
|
||
|
||
### 导出路径说明
|
||
|
||
| 标注类型 | 导出产物 | 条件 |
|
||
|---------|---------|------|
|
||
| cuboid (3D) | `labels/quaternion_json/*.json` | CVAT cuboid 标注 |
|
||
| rectanglelabels (2D) | `labels/quaternion_json/*.json` | 2D 矩形自动桥接 |
|
||
| YOLO HBB | `labels/yolo-hbb/*.txt` | 无 cuboid 时使用 `export_adas_yolo` |
|
||
|
||
## CLI
|
||
|
||
```bash
|
||
# cuboid 导出（平台 Job）
|
||
PYTHONPATH=platform python3 -c "from as_platform.labeling.export_cuboid_batch import export_batch; ..."
|
||
|
||
# YOLO HBB 导出（平台 Job，无 cuboid 时自动选择）
|
||
# 由 jobs/runner.py 调用 datasets/adas/scripts/export_ls_to_yolo.py
|
||
|
||
# 3D 拟合
|
||
PYTHONPATH=platform python3 -c "from as_platform.labeling.fit_cuboid_batch import fit_batch; ..."
|
||
|
||
# 入包
|
||
python as.py build adas cuboid_7cls --batch val_front6mm_pilot --pack adas_v1
|
||
```
|
||
|
||
## Smoke
|
||
|
||
```bash
|
||
bash scripts/smoke_adas_promote.sh
|
||
```
|
||
|
||
## 数据目录
|
||
|
||
ADAS 前向数据在 **数据目录** → **前向** 视图中展示，按 pack → batch 层级显示图片数、标注数和类别分布。
|