facial-landmarks-35-adas-0002

This is a custom-architecture convolutional neural network for 35 facial landmarks estimation.

[Left Eye] **p0, p1**: corners of the eye, located on the boundary of the eyeball and the eyelid.

[Right Eye] **p2, p3**: corners of the eye, located on the boundary of the eyeball and the eyelid.

[Nose] **p4**: nose-tip point; **p5**: lowest point of the nasal septum; **p6, p7**: right-bottom and left-bottom of the nose wing.

[Mouth] **p8, p9**: mouth corners on the outer boundary of the lip; **p10, p11**: center points along the outer boundary of the lip.

[Left Eyebrow] **p12**: starting point of the upper boundary of the eyebrow; **p13**: mid-point of the upper arc of the eyebrow; **p14**: ending point of the upper boundary of the eyebrow.

[Right Eyebrow] **p15**: starting point of the upper boundary of the eyebrow; **p16**: mid-point of the upper arc of the eyebrow; **p17**: ending point of the upper boundary of the eyebrow.

[Face Contour] **p26**: chin center; **p18, p34**: upper points of the face contour aligned with the outer corners of the eyes; **p19~p25**: boundary points, evenly distributed along the curve p18-p26; **p27~p33**: boundary points, evenly distributed along the curve p26-p34.

Metric | Value |
---|---|

GFlops | 0.042 |

MParams | 4.595 |

Source framework | Caffe* |

A 1000-sample random subset of a large internal dataset containing images of 300 people with different facial expressions.

The quality of landmarks' positions prediction is evaluated through the use of Normed Error (NE). The error for the i^{th} sample has the form:

where N is the number of landmarks, *p*-hat and *p* are, correspondingly, the prediction and ground truth vectors of the k^{th} landmark of the i^{th} sample, and d_{i} is the interocular distance for the i^{th} sample.

Dataset | Mean NE | 90^{th} Percentile NE | Standard deviation of NE |
---|---|---|---|

Internal dataset | 0.106 | 0.143 | 0.038 |

- Blob in the format [BxCxHxW] where:
- B - batch size
- C - number of channels
- H - image height
- W - image width

with the name `data`

and the shape [1x3x60x60].

The net outputs a blob with the shape: [1, 70], containing row-vector of 70 floating point values for 35 landmarks' normed coordinates in the form (x0, y0, x1, y1, ..., x34, y34).

Output layer name in Inference Engine format:

`align_fc3`

Output layer name in Caffe* format:

`align_fc3`

[*] Other names and brands may be claimed as the property of others.