What Facial Landmarks Does RetinaFace Detect?
RetinaFace detects five key facial landmarks for every face it locates in an image. These five points provide sufficient geometric information for face alignment, the most critical preprocessing step in face recognition pipelines.
The Five Facial Landmarks
- Left eye center — the center of the left eye pupil region
- Right eye center — the center of the right eye pupil region
- Nose tip — the lowest visible point of the nose bridge
- Left mouth corner — the left commissure of the lips
- Right mouth corner — the right commissure of the lips
Each landmark is expressed as an (x, y) pixel coordinate relative to the original input image dimensions.
Why These Five Points?
Five-point landmark sets have become the standard in face recognition research because they strike the ideal balance between geometric informativeness and annotation cost. The pair of eye centers establishes the inter-ocular line which defines the roll angle of the head. The nose tip provides a reference for pitch, and the two mouth corners complete the constraint for yaw.
How Are Landmarks Used for Face Alignment?
Face alignment is the process of warping a detected face into a canonical standardized orientation. Given the five predicted landmark coordinates, a similarity transformation is computed between the predicted points and a set of reference positions in a template coordinate system, producing an aligned face crop typically 112x112 pixels.
Steps in the Alignment Process
- Run RetinaFace to get the five landmark coordinates for each face
- Define a set of five reference template points
- Compute the optimal similarity transformation using least squares
- Apply the transformation to the original image using bilinear interpolation
- Crop the output to the standard face crop size (112x112)
Conclusion
RetinaFace detects five strategically chosen facial landmarks that provide everything needed for accurate face alignment and a wide range of downstream analysis tasks. Understanding how these landmarks are predicted and used is essential for building robust face processing pipelines.
