From 1a01ddf955925eee8694a7bc0609b3f1ba09f13b Mon Sep 17 00:00:00 2001 From: Neucrack Date: Fri, 21 Jun 2024 19:20:52 +0800 Subject: [PATCH] add yolov8 face demo --- docs/doc/en/vision/face_detection.md | 45 +++++++++++++++------ docs/doc/zh/vision/face_detection.md | 37 +++++++++++++---- examples/vision/ai_vision/nn_yolov8_face.py | 16 ++++++++ 3 files changed, 78 insertions(+), 20 deletions(-) create mode 100644 examples/vision/ai_vision/nn_yolov8_face.py diff --git a/docs/doc/en/vision/face_detection.md b/docs/doc/en/vision/face_detection.md index cca1ae7..e6261e0 100644 --- a/docs/doc/en/vision/face_detection.md +++ b/docs/doc/en/vision/face_detection.md @@ -1,29 +1,44 @@ --- -title: MaixPy Face Detection and Key Points Detection +title: MaixPy Face Detection and Keypoint Detection --- ## Introduction -Face detection can be used in many places, such as providing the step of face detection for face recognition, or applications related to face tracking, and more. +Face detection can be applied in many scenarios, such as providing the face detection step for face recognition, or for face tracking applications, etc. -The face detection provided here can not only detect faces but also detect 5 key points, including two eyes, one nose, and the two corners of a mouth. +The face detection provided here can not only detect faces but also detect 5 key points, including two eyes, one nose, and two corners of the mouth. ![face detection](../../assets/face_detection.jpg) ## Using Face Detection in MaixPy -MaixPy officially provides two face detection models, sourced from the open projects [face detector 1MB with landmark](https://github.com/biubug6/Face-Detector-1MB-with-landmark) and [Retinaface](https://github.com/biubug6/Pytorch_Retinaface). +MaixPy officially provides three face detection models from the open-source projects [Face Detector 1MB with landmark](https://github.com/biubug6/Face-Detector-1MB-with-landmark), [Retinaface](https://github.com/biubug6/Pytorch_Retinaface), and [YOLOv8-face](https://github.com/derronqi/yolov8-face). -To use them, first download a model, either one as there's not much difference between them: -* [face detector 1MB with landmark](https://maixhub.com/model/zoo/377) -* [Retinaface](https://maixhub.com/model/zoo/378) +All three models can be used. `YOLOv8-face` performs better but is slightly slower, so you can choose based on your testing. + +Using `YOLOv8-face` (requires MaixPy version >= 4.3.8): + +```python +from maix import camera, display, image, nn, app -Then copy the model file to your device, see [Using MaixVision](../basic/maixvision.md) for how to copy. -> The default image contains a file that can be used directly; if not available, you must download it yourself. The downloaded zip package contains multiple resolutions to choose from; the higher the resolution, the more precise but also more time-consuming. +detector = nn.YOLOv8(model="/root/models/yolov8n_face.mud") -Next, run the code. The following line of commented code is for loading the `Retinaface` model, choose which line of code to use based on the model you downloaded. +cam = camera.Camera(detector.input_width(), detector.input_height(), detector.input_format()) +dis = display.Display() + +while not app.need_exit(): + img = cam.read() + objs = detector.detect(img, conf_th = 0.5, iou_th = 0.45, keypoint_th = 0.5) + for obj in objs: + img.draw_rect(obj.x, obj.y, obj.w, obj.h, color = image.COLOR_RED) + msg = f'{detector.labels[obj.class_id]}: {obj.score:.2f}' + img.draw_string(obj.x, obj.y, msg, color = image.COLOR_RED) + detector.draw_pose(img, obj.points, 2, image.COLOR_RED) + dis.show(img) +``` -> To use this function, MaixPy must >= 4.1.4. +For the other two models: +Here, a line of commented-out code is used to load the `Retinaface` model. Choose which line of code to use based on the model you download. ```python from maix import camera, display, image, nn, app @@ -43,5 +58,11 @@ while not app.need_exit(): radius = math.ceil(obj.w / 10) img.draw_keypoints(obj.points, image.COLOR_RED, size = radius if radius < 5 else 4) dis.show(img) - ``` + +## Model Downloads and Other Resolution Models + +Download the models; the compressed package contains multiple resolutions to choose from. Higher resolution models are more accurate but take longer to process: +* [Face Detector 1MB with landmark](https://maixhub.com/model/zoo/377) +* [Retinaface](https://maixhub.com/model/zoo/378) +* [YOLOv8-face](https://maixhub.com/model/zoo/407) diff --git a/docs/doc/zh/vision/face_detection.md b/docs/doc/zh/vision/face_detection.md index 2fa772b..4068cb7 100644 --- a/docs/doc/zh/vision/face_detection.md +++ b/docs/doc/zh/vision/face_detection.md @@ -13,19 +13,33 @@ title: MaixPy 人脸检测和关键点检测 ## MaixPy 中使用人脸检测 -MaixPy 官方提供了两种人脸检测模型，分别来自开源项目 [face detector 1MB with landmark](https://github.com/biubug6/Face-Detector-1MB-with-landmark) 和 [Retinafate](https://github.com/biubug6/Pytorch_Retinaface)。 +MaixPy 官方提供了三种人脸检测模型，分别来自开源项目 [face detector 1MB with landmark](https://github.com/biubug6/Face-Detector-1MB-with-landmark) 和 [Retinafate](https://github.com/biubug6/Pytorch_Retinaface) 以及 [YOLOv8-face](https://github.com/derronqi/yolov8-face)。 -要使用需要先下载模型，选择一个即可，两者区别不大： -* [face detector 1MB with landmark](https://maixhub.com/model/zoo/377) -* [Retinafate](https://maixhub.com/model/zoo/378) +这三种模型都可以用，`YOLOv8-face` 效果比较好但是速度略微慢一些，可以自己实际测试选择使用。 + +使用`YOLOv8-face`：（需要 MaixPy 版本 >= 4.3.8） -然后拷贝模型文件到设备，拷贝方法见 [MaixVision 使用](../basic/maixvision.md)。 -> 默认镜像里面有一个文件，可以直接使用，如果没有则需要你自己下载，而且下载的压缩包里面有多个分辨率可以选择，分辨率越高越精准但耗时更长 +```python +from maix import camera, display, image, nn, app -然后执行代码，这里有一行被注释了代码是加载`Retinafae`模型，根据你下载的模型选择使用哪一行代码 +detector = nn.YOLOv8(model="/root/models/yolov8n_face.mud") -> 本功能需要 MaixPy >= 4.1.4 才能使用 +cam = camera.Camera(detector.input_width(), detector.input_height(), detector.input_format()) +dis = display.Display() + +while not app.need_exit(): + img = cam.read() + objs = detector.detect(img, conf_th = 0.5, iou_th = 0.45, keypoint_th = 0.5) + for obj in objs: + img.draw_rect(obj.x, obj.y, obj.w, obj.h, color = image.COLOR_RED) + msg = f'{detector.labels[obj.class_id]}: {obj.score:.2f}' + img.draw_string(obj.x, obj.y, msg, color = image.COLOR_RED) + detector.draw_pose(img, obj.points, 2, image.COLOR_RED) + dis.show(img) +``` +另外两种模型使用方法： +这里有一行被注释了代码是加载`Retinafae`模型，根据你下载的模型选择使用哪一行代码 ```python from maix import camera, display, image, nn, app @@ -49,6 +63,13 @@ while not app.need_exit(): ``` +## 模型下载和其它分辨率模型 + +下载模型，下载的压缩包里面有多个分辨率可以选择，分辨率越高越精准但耗时更长： +* [face detector 1MB with landmark](https://maixhub.com/model/zoo/377) +* [Retinafate](https://maixhub.com/model/zoo/378) +* [YOLOv8-face](https://maixhub.com/model/zoo/407) + diff --git a/examples/vision/ai_vision/nn_yolov8_face.py b/examples/vision/ai_vision/nn_yolov8_face.py new file mode 100644 index 0000000..d64d9e4 --- /dev/null +++ b/examples/vision/ai_vision/nn_yolov8_face.py @@ -0,0 +1,16 @@ +from maix import camera, display, image, nn, app + +detector = nn.YOLOv8(model="/root/models/yolov8n_face.mud") + +cam = camera.Camera(detector.input_width(), detector.input_height(), detector.input_format()) +dis = display.Display() + +while not app.need_exit(): + img = cam.read() + objs = detector.detect(img, conf_th = 0.5, iou_th = 0.45, keypoint_th = 0.5) + for obj in objs: + img.draw_rect(obj.x, obj.y, obj.w, obj.h, color = image.COLOR_RED) + msg = f'{detector.labels[obj.class_id]}: {obj.score:.2f}' + img.draw_string(obj.x, obj.y, msg, color = image.COLOR_RED) + detector.draw_pose(img, obj.points, 2, image.COLOR_RED) + dis.show(img)