File size: 3,749 Bytes
a097504
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15240f0
a097504
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15240f0
 
 
 
 
 
 
a097504
 
 
 
 
 
 
 
15240f0
 
 
 
a097504
 
 
 
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
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
---
license: apache-2.0
library_name: onnx
tags:
  - computer-vision
  - image-classification
  - face-analysis
  - age-estimation
  - gender-classification
  - onnx
  - cpu
pipeline_tag: image-classification
---

# FastFace

FastFace is a CPU-efficient face age and gender model family. Phase 1 freezes three variants:

| Variant | Purpose | Input | Primary artifact |
| --- | --- | --- | --- |
| `fastface-large-128` | Recommended accuracy/throughput student | aligned RGB face crop, 128x128 | `models/fastface-large-128/model_fp32.onnx` |
| `fastface-small-112` | Highest-throughput student | aligned RGB face crop, 112x112 | `models/fastface-small-112/model_fp32.onnx` |
| `fastface-teacher-v2s-128` | Audit/teacher model, not CPU default | aligned RGB face crop, 128x128 | `models/fastface-teacher-v2s-128/model_fp32.onnx` |
| `fastfacedetector-retinaface-mnetv1-960` | Full-image face detector and 5-point alignment stage | raw BGR/RGB image | `models/fastfacedetector-retinaface-mnetv1-960/fastfacedetector_retinaface_mobilenetv1_050_whole960_epoch34.onnx` |

The released task is gender classification plus numeric age estimation. Race prediction is intentionally out of scope.

## Training Data

Training used aligned face crops and manifests built from public/research datasets available in the local training workspace:

- FairFace train/validation labels.
- UTKFace aligned face images.
- IMDB-clean derived from the IMDB-WIKI family after quality filtering.
- Lagenda-hosted face-age data was explored but is not part of the frozen phase-1 student release.

See `data_provenance.md` and `technical_report.md` in this repository for the exact source policy, manifest contract, and limitations.

## Metrics

Validation gender balanced accuracy on the mixed public validation set:

| Model | Mixed GBA | FairFace GBA | IMDB-clean GBA | UTKFace GBA |
| --- | ---: | ---: | ---: | ---: |
| `fastface-teacher-v2s-128` | 0.98605 | 0.94386 | 0.99138 | 0.95424 |
| `fastface-large-128` | 0.97929 | 0.92877 | 0.98548 | 0.95017 |
| `fastface-small-112` | 0.96800 | 0.90562 | 0.97542 | 0.94101 |

In a 24,333-sample comparison set, gender balanced accuracy was:

- `fastface-teacher-v2s-128`: 0.96638
- `fastface-large-128`: 0.95618
- public `fairface-onnx`: 0.94658
- `fastface-small-112`: 0.94059

Public FairFace-ONNX vs `fastface-large-128` disagreed on 1,301 samples. Against the available public labels, `fastface-large-128` was correct on 762 of those and public FairFace-ONNX was correct on 539.

The detector candidate was evaluated on full WIDER FACE validation:

| Detector | Precision | Recall | F1 | Seconds/Image |
| --- | ---: | ---: | ---: | ---: |
| `fastfacedetector-retinaface-mnetv1-960` | 0.87832 | 0.48082 | 0.62144 | 0.02088 |
| UniFace RetinaFace MNetV2 baseline | 0.87753 | 0.44552 | 0.59099 | 0.02699 |

## Intended Use

- High-throughput CPU inference on already-detected/aligned face crops.
- Gender and age signals for product analytics or moderation-assist workflows where uncertainty and bias are handled upstream/downstream.
- Teacher-assisted auditing and future distillation.

## Limitations

- The age/gender models do not detect faces; they expect aligned face crops.
- The detector candidate provides the raw-image face detection and 5-point
  alignment stage, but its alignment benchmark uses UniFace RetinaFace MNetV2 as
  the comparison baseline rather than human landmark annotations.
- Age labels are noisy across public face-age datasets, so age should be treated as an estimate.
- Gender labels follow dataset annotations and can encode social and labeling bias.
- Race/ethnicity classification is not provided.
- Metrics reflect the assembled validation manifests, not universal real-world performance.