Jm/update

.gitignore

@@ -45,12 +45,15 @@ build/
 dist/
 .vscode/
 coco-*/
+dota-*/
 imagenet-1k*/
 tmp/
+DOTAv1/
 val2017/
 .omc
 .bak
 test_bash_script
+venv*
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

AGENTS.md

@@ -18,9 +18,12 @@ Markdown guides, example scripts, or both.
 - `compilation/`: Tutorials and helper scripts for qbcompiler workflows
 - `runtime/`: Tutorials and helper scripts for qbruntime workflows
 - `assets/`: Shared images used by the docs
+- `compilation/oriented_bounding_boxes/`: `YOLO11m-obb` compilation tutorial for DOTAv1-style OBB models
+- `runtime/python/oriented_bounding_boxes/`: Self-contained MXQ runtime tutorial for OBB inference and visualization
 - `compilation/*/README.md`, `runtime/*/README.md`: English task-specific guides
 - `compilation/*/README.KR.md`, `runtime/*/README.KR.md`: Korean counterparts when available
 - `compilation/*/*.py`, `runtime/*/*.py`: Standalone example scripts used by the guides
+- `runtime/python/*/{postprocess,visualize,utils}.py`: Local helper modules that intentionally stay close to each tutorial
 - `*/**/requirements.txt`: Per-tutorial dependency hints for heavier examples such as LLM, STT,
   and VLM flows
 
@@ -31,6 +34,8 @@ Markdown guides, example scripts, or both.
   describes.
 - Prefer small, explicit scripts over shared abstractions unless the repo already has a local
   utility module for that example.
+- For runtime vision tutorials, prefer local preprocessing, postprocessing, dataset-label, and
+  visualization helpers when that keeps the example easier to follow end to end.
 - Assume users follow the docs step by step on Linux or in Docker, often with Mobilint NPU access.
 
 ## Documentation Rules
@@ -92,6 +97,9 @@ Markdown guides, example scripts, or both.
   downloads, or external packages, document that explicitly in the README.
 - Keep preprocessing, postprocessing, and model-loading logic close to the example unless there is
   already a local helper module such as `imagenet.py`, `coco.py`, `utils.py`, or `visualize.py`.
+- For OBB examples, keep the row format and dataset assumptions explicit in code and docs, for
+  example `cx, cy, w, h, conf, cls, angle` with DOTAv1 labels when the tutorial is tied to
+  `YOLO11m-obb`.
 
 ### Dependency Awareness
 

SKILL.md

@@ -0,0 +1,92 @@
+---
+description: Reusable repository skill for making aligned documentation and example-script changes in the Mobilint SDK tutorial repo.
+paths:
+  - "**"
+---
+
+# Mobilint SDK Tutorial Skill
+
+## When to Use This Skill
+
+Use this skill for changes anywhere in this repository, especially when the task involves:
+
+- Tutorial docs under `README.md`, `README.KR.md`, `compilation/`, or `runtime/`
+- Standalone example scripts that accompany a specific tutorial
+- Keeping script arguments, filenames, and README commands synchronized
+- Updating bilingual documentation where English and Korean versions coexist
+- Validation planning for workflows that depend on `qbcompiler`, `qbruntime`, Docker, NPU devices,
+  gated datasets, or large model downloads
+
+## Purpose
+
+Use this repo as a documentation-first tutorial workspace. Changes should keep example scripts,
+README commands, and expected artifacts synchronized within each tutorial directory.
+
+## First Reads
+
+Start with the smallest set of files that anchor the task:
+
+- `README.md` and `README.KR.md` for the top-level product framing
+- The nearest `compilation/<task>/README.md` or `runtime/<task>/README.md`
+- The adjacent script files used by that tutorial
+- `pyproject.toml` for Python version and Ruff settings
+- `git status --short` before editing so you do not overwrite unrelated user work
+
+If the touched tutorial has a Korean counterpart, open `README.KR.md` early so structure changes
+can stay aligned.
+
+## Current Repo State
+
+- The repo contains paired compilation and runtime tutorials under `compilation/` and `runtime/`.
+- Many tutorials have both `README.md` and `README.KR.md`; when both exist, keep them aligned.
+- Runtime Python tutorials commonly use small local helper modules such as `utils.py`,
+  `postprocess.py`, `visualize.py`, and dataset label files like `coco.py` or `dota.py`.
+- `compilation/oriented_bounding_boxes/` contains the `YOLO11m-obb` compilation flow that produces
+  `yolo11m-obb.mxq`.
+- `runtime/python/oriented_bounding_boxes/` is a self-contained OBB runtime tutorial that expects
+  `../../../compilation/oriented_bounding_boxes/yolo11m-obb.mxq`, uses `1024x1024` letterbox
+  preprocessing, keeps runtime input as `uint8`, decodes DOTA OBB rows as
+  `cx, cy, w, h, conf, cls, angle`, and renders rotated polygons.
+
+## Repo Map
+
+- `compilation/README.md`: Compiler setup, Docker workflow, and qbcompiler installation guidance
+- `runtime/README.md`: Runtime setup, driver/library installation, and NPU assumptions
+- `compilation/*`: Model-specific compilation walkthroughs
+- `runtime/python/*`: Python runtime walkthroughs
+- `runtime/cpp/*`: C++ runtime walkthroughs where present
+- `assets/`: Shared diagrams and logo assets
+- Per-directory helper modules such as `imagenet.py`, `coco.py`, `dota.py`, `postprocess.py`,
+  `utils.py`, and `visualize.py`: Local helpers for one tutorial family
+
+## Working Rules
+
+- Treat each tutorial as self-contained unless there is already a local shared helper in that area.
+- When you change script defaults or CLI flags, update the README in the same directory immediately.
+- Prefer concrete filenames and copy-pasteable commands.
+- Keep bilingual tutorial structure synchronized even when the prose is not a literal translation.
+- Prefer direct, readable scripts that mirror the tutorial text instead of introducing shared
+  library-style abstractions.
+- Keep README sections ordered around user workflow: prerequisites, preparation, execution, output.
+- Reflect external constraints explicitly when examples depend on Mobilint proprietary wheels,
+  Mobilint NPU devices, Docker images, Hugging Face authentication, or large downloads.
+- For hardware-dependent examples, do static validation when full execution is not available and
+  state that limit clearly.
+
+## Validation Defaults
+
+For touched Python tutorial files, prefer:
+
+```bash
+ruff check path/to/file.py
+ruff format path/to/file.py
+python -m py_compile path/to/file.py
+```
+
+For a small tutorial directory with several edited scripts:
+
+```bash
+python -m compileall path/to/tutorial_dir
+```
+
+For touched docs, verify that links, paths, commands, and default filenames match the local scripts.

compilation/face_detection/README.KR.md

@@ -124,6 +124,7 @@ python model_compile.py --onnx-path ./yolov12m-face.onnx --calib-data-path ./wid
 ```
 
 이 예제는 `prepare_model.py`의 출력 파일명과 맞추기 위해 경로를 명시적으로 전달합니다.
+ARIES 튜토리얼 스크립트는 내부적으로 `target_device="aries2"`로 컴파일하므로, 이 명령은 기본적으로 ARIES2를 대상으로 합니다.
 
 **수행 작업:**
 
@@ -143,4 +144,16 @@ python model_compile.py --onnx-path ./yolov12m-face.onnx --calib-data-path ./wid
 - `yolov12m-face.mxq`
 - `yolov12m-face.mblt`
 
+### REGULUS
+
+REGULUS는 `inference_scheme="single"`만 지원합니다. `model_compile_regulus.py`를 `--target-device`와 함께 사용하세요.
+
+> **참고**: face detection 은 **regulus2 에서만** 지원됩니다. `yolo-face` 프로젝트가 YOLOv8 기반 face weight 를 제공하지 않고 regulus1 은 YOLOv9 이하만 지원하므로, YOLOv12 face 모델은 regulus1 을 대상으로 할 수 없습니다.
+
+```bash
+python model_compile_regulus.py --onnx-path ./yolov12m-face.onnx --calib-data-path ./widerface-selected --save-path ./yolov12m-face.mxq --target-device regulus2
+```
+
+현재 디렉토리에 `yolov12m-face.mxq` (및 중간 산출물 `yolov12m-face.mblt`) 가 생성됩니다.
+
 명령이 끝나면 [../../runtime/python/face_detection/README.KR.md](../../runtime/python/face_detection/README.KR.md)에서 컴파일된 모델로 추론을 실행할 수 있습니다.

compilation/face_detection/README.md

@@ -124,6 +124,7 @@ python model_compile.py --onnx-path ./yolov12m-face.onnx --calib-data-path ./wid
 ```
 
 This example passes the paths explicitly so the command matches the output from `prepare_model.py`.
+The ARIES tutorial script now compiles with `target_device="aries2"` internally, so this command targets ARIES2 by default.
 
 **What this does:**
 
@@ -143,4 +144,16 @@ This example passes the paths explicitly so the command matches the output from
 - `yolov12m-face.mxq`
 - `yolov12m-face.mblt`
 
+### REGULUS
+
+REGULUS only supports `inference_scheme="single"`. Use `model_compile_regulus.py` with `--target-device`.
+
+> **Note**: Face detection is supported on **regulus2 only**. The `yolo-face` project does not ship a YOLOv8-based face weight, and regulus1 supports only YOLOv9 and earlier, so the YOLOv12 face model cannot target regulus1.
+
+```bash
+python model_compile_regulus.py --onnx-path ./yolov12m-face.onnx --calib-data-path ./widerface-selected --save-path ./yolov12m-face.mxq --target-device regulus2
+```
+
+This produces `yolov12m-face.mxq` (and the intermediate `yolov12m-face.mblt`) in the current directory.
+
 After the command finishes, continue to [../../runtime/python/face_detection/README.md](../../runtime/python/face_detection/README.md) to run inference with the compiled model.

compilation/face_detection/model_compile.py

@@ -50,6 +50,7 @@
     )
     mxq_compile(
         model=args.onnx_path,
+        target_device="aries2",
         calib_data_path=args.calib_data_path,
         save_subgraph_type=2,  # save mblt file before quantization
         output_subgraph_path=args.onnx_path.replace(".onnx", ".mblt"),