How to get the results of landmarks with 68 points as dlib

[zoeleesss]

Description

How to get the results of landmarks with 68 points as dlib. I don't wanna use import dlib; etc.. . Cause it takes too much time. When i was trying to get landmarks, i used api.face_landmarks(image). But what i got is 72 points. And i tried to manually changed the results to 68 points according to the results used by dlib.

What i did was transforming 72 points to 68 points:

def compatibleToDlib(arr):
    results = []
    # 0 - 54
    for i1 in range(55):
        results.append(arr[i1])

    # 55,56,57,58,59  <-  arr[61-65]
    for i2 in range(61,66):
        results.append(arr[i2])

    # 60 <- arr[59]
    results.append(arr[59])

    # 61,62,63 <- arr[58-56]
    for i3 in range(0,3):
        results.append(arr[-i3+58])

    # 64 <- arr[55]
    results.append(arr[55])

    # 65,66,67 <- arr[70,69,68]
    for i4 in range(0,3):
        results.append(arr[-i4+70])

    # print(len(results))
    return results

But it didnt work tho. Cause the project that i needed to work with is deepfakes and tried the conversion part afterwards, but it says wrong aligened. shapes (2,51) and (55,2) not aligned: 51 (dim 1) != 55 (dim 0) And i tried to replace the landmarks by the landmarks created by dlib. And it works.

So what i want to know is how to get the same aligned results of 68-point landmarks as dlib by only using this repo - face_recognition ?

Thanks so much!!

[frederickszk]

Hello! I also encounter the same problem and resolve it by reading the codes in face_recognition/api.py . In L177~L192, it invokes the Dlib landmark detector and obtains the 68 landmarks at first. Afterward, it processes them into 72 landmarks. Therefore I simply extract 68 landmarks by:

from face_recognition.api import _raw_face_landmarks
...
landmarks = _raw_face_landmarks(img_frame)
face_landmark_list = [[(p.x, p.y) for p in landmark.parts()] for landmark in landmarks]

However, from my evaluation results, I found that it's about 6x slower than directly using Dlib=19.23.1.