有没有办法通过 Mediapipe 将 468 个地标映射/转换为 68 个地标？

发布于 2025-01-10 15:45:19 字数 158 浏览 0 评论 0原文

我试图将地面真实面部标志（68 个标志）与 Mediapipe 标志检测（468 个标志）进行比较。为了做到这一点，我认为我需要以某种方式将 468 个地标映射到 68 个地标。我可能的解决方案是手动查找最接近 68 个地标中每个地标的索引并输出它们。但我不确定这里的准确性。有人可以在这方面帮助我吗？

原文

分享到QQ

分享到微博

如果你对这篇内容有疑问，欢迎到本站社区发帖提问参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。

发布评论

需要登录才能够评论，你可以免费注册一个本站的账号。

执着的年纪 2025-01-17 15:45:19

我不是该主题的专家，但我认为没有直接的方法来进行转换，这是因为一侧到另一侧的映射不同。

对此，我采取了与您提到的相同的想法，并且提取了最接近的点。

这是我添加到 mediapipe 建议在 dlib 中使用的代码中的代码。考虑到许多点彼此重合，最终结果并没有那么错误。

在 MediaPipe Face Mesh 代码示例的开头，您应该添加一个列表，该列表是与 dlib 68 地标匹配的点的选择：

import cv2
import mediapipe as mp
mp_drawing = mp.solutions.drawing_utils
mp_drawing_styles = mp.solutions.drawing_styles
mp_face_mesh = mp.solutions.face_mesh

#New Add
landmark_points_68 = [162,234,93,58,172,136,149,148,152,377,378,365,397,288,323,454,389,71,63,105,66,107,336,
                  296,334,293,301,168,197,5,4,75,97,2,326,305,33,160,158,133,153,144,362,385,387,263,373,
                  380,61,39,37,0,267,269,291,405,314,17,84,181,78,82,13,312,308,317,14,87]

在 MediaPipe Face Mesh 代码示例中查找该行：

for face_landmarks in results.multi_face_landmarks:

然后添加以下内容：

        landmarks_extracted = []
        for index in landmark_points_68:
            x = int(face_landmarks.landmark[index].x * width)
            y = int(face_landmarks.landmark[index].y * height)
            landmarks_extracted.append((x, y))

现在该列表（提取的地标）你可以在你的代码中使用它

I'm not an expert on the subject, but I think there is no direct way to do a conversion, this is because the mapping is different from one side to the other.

To this I have taken the same idea that you mention and I have extracted the closest points.

So this is the code that I have added to the code that mediapipe proposes for my use in dlib. The final result is not so wrong considering that many points are coincident with each other.

At the beginning of the MediaPipe Face Mesh code example you should add a list which is a selection of points that match dlib 68 landmarks:

import cv2
import mediapipe as mp
mp_drawing = mp.solutions.drawing_utils
mp_drawing_styles = mp.solutions.drawing_styles
mp_face_mesh = mp.solutions.face_mesh

#New Add
landmark_points_68 = [162,234,93,58,172,136,149,148,152,377,378,365,397,288,323,454,389,71,63,105,66,107,336,
                  296,334,293,301,168,197,5,4,75,97,2,326,305,33,160,158,133,153,144,362,385,387,263,373,
                  380,61,39,37,0,267,269,291,405,314,17,84,181,78,82,13,312,308,317,14,87]

In the MediaPipe Face Mesh code example look for the line:

for face_landmarks in results.multi_face_landmarks:

then add the following:

        landmarks_extracted = []
        for index in landmark_points_68:
            x = int(face_landmarks.landmark[index].x * width)
            y = int(face_landmarks.landmark[index].y * height)
            landmarks_extracted.append((x, y))

now that list (landmarks extracted) you can use it in your code

回复收藏 0 原文

西瑶 2025-01-17 15:45:19

这段代码可以做到：
https://github.com/PeizhiYan/Mediapipe_2_Dlib_Landmarks/

此代码将 Mediapipe 的 478 个密集面部标志映射到Dlib 的 68 个稀疏面部标志通过定义对应关系，其中每个 Dlib 标志索引对应于一个或两个 Mediapipe 索引，必要时对坐标进行平均。函数convert_landmarks_mediapipe_to_dlib 使用此预定义映射将 Mediapipe 地标数组转换为 Dlib 地标。

回复收藏 0 原文

~没有更多了~