face_id 0.2.0

Face detection (bounding boxes, eyes/nose/mouth markers), and facial recognition (embeddings).
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39

# Face ID

Face detection & recognition crate.

![img.png](.github/readme_img.png)

## Face detection

This crate uses SCRDF face detection models. The following models are available:

> The naming convention for the ONNX models indicates the computational complexity (measured in FLOPs) and whether the
> model includes 5 facial keypoints predictions in addition to standard bounding boxes.

|      Name       | Easy  | Medium | Hard  | FLOPs | Params(M) | Infer(ms) | BBox | Facial Keypoints |
|:---------------:|-------|--------|-------|-------|-----------|-----------|:-----|:-----------------|
|    500m.onnx    | 90.57 | 88.12  | 68.51 | 500M  | 0.57      | 3.6       | ✅    | ❌                |
|     1g.onnx     | 92.38 | 90.57  | 74.80 | 1G    | 0.64      | 4.1       | ✅    | ❌                |
|    34g.onnx     | 96.06 | 94.92  | 85.29 | 34G   | 9.80      | 11.7      | ✅    | ❌                |
| 2.5g_bnkps.onnx | 93.80 | 92.02  | 77.13 | 2.5G  | 0.82      | 4.3       | ✅    | ✅                |
| 10g_bnkps.onnx  | 95.40 | 94.01  | 82.80 | 10G   | 4.23      | 5.0       | ✅    | ✅                |
| 34g_gnkps.onnx  | ?     | ?      | ?     | 34G   | ?         | ?         | ✅    | ✅                |

### Keypoints (`kps`) and Normalization Types (`bn` vs `gn`)

- **`kps`**: Denotes models that output 5 facial landmarks (keypoints) in addition to the standard bounding boxes.
- **`bnkps`**: Models trained using **Batch Normalization (BN)**. These often have lower false-positive rates and high
  recall on general datasets. However, they occasionally struggle with producing accurate landmarks for faces that are
  rotated past 90 degrees or are unusually large.
- **`gnkps`**: Models trained using **Group Normalization (GN)**. These variants (e.g., `34g_gnkps` or `10g_gnkps`) were
  explicitly developed to fix issues with very large faces that the `bnkps` models exhibited. While they improve
  landmark quality on large or rotated faces, they might have slightly lower general recall than `bnkps`.

## Face recognition (embeddings)

Use ArcFace ... explain more here. it needs to get aligned faces where the mouth, nose, eyes are in exact spots in an
112x112 image.
this is handled by this crate.

you can use hf model id: public-data/insightface with filename "models/buffalo_l/w600k_r50.onnx" to embed faces.