File size: 5,504 Bytes
289dffb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
---
license: other
task_categories:
- image-classification
- zero-shot-image-classification
tags:
- face-recognition
- face-verification
- face-identification
- ijb-a
- janus
- template-based
pretty_name: IJB-A HF-ready
size_categories:
- 10K<n<100K
---

# IJB-A HF-ready

This repo packages the IARPA Janus Benchmark-A (IJB-A) face recognition dataset
in its CleanData layout, plus the official 10-split 1:1 verification and 1:N
identification protocols.

Unlike LFW-style benchmarks, IJB-A is **template-based**: each subject is
represented by a *template* aggregating multiple still images and/or video
frames. Protocol CSVs map every (template, file) row to a face annotation with
bounding box, landmarks, and demographic attributes.

## Layout

```
ijba/
├── README.md
├── files.csv
├── img/                              # 5396 still images (flat)
├── frame.tar                         # 20369 video frames packed into a single tar (~817 MB)
├── img.txt                           # upstream bbox annotations subset (still images)
├── frame.txt                         # upstream bbox annotations subset (video frames)
└── protocols/
    ├── IJB-A_11_sets/split{1..10}/    # 1:1 verification protocol inputs
    │   ├── train_<n>.csv
    │   ├── verify_metadata_<n>.csv
    │   └── verify_comparisons_<n>.csv
    ├── IJB-A_11_output/split{1..10}/  # baseline 1:1 outputs (.matches)
    ├── IJB-A_1N_sets/split{1..10}/    # 1:N identification protocol inputs
    │   ├── train_<n>.csv
    │   ├── search_gallery_<n>.csv
    │   └── search_probe_<n>.csv
    └── IJB-A_1N_output/split{1..10}/  # baseline 1:N outputs (.candidate_lists)
```

## Protocol CSV columns

`train_*.csv`, `verify_metadata_*.csv`, `search_gallery_*.csv`, `search_probe_*.csv`
all share the same schema:

`TEMPLATE_ID, SUBJECT_ID, FILE, MEDIA_ID, SIGHTING_ID, FRAME, FACE_X, FACE_Y,
FACE_WIDTH, FACE_HEIGHT, RIGHT_EYE_X, RIGHT_EYE_Y, LEFT_EYE_X, LEFT_EYE_Y,
NOSE_BASE_X, NOSE_BASE_Y, FACE_YAW, FOREHEAD_VISIBLE, EYES_VISIBLE,
NOSE_MOUTH_VISIBLE, INDOOR, GENDER, SKIN_TONE, AGE, FACIAL_HAIR`

The `FILE` column resolves relative to the dataset root, e.g. `img/8565.jpg` or
`frame/28065_00000.png`.

`verify_comparisons_*.csv` is a header-less file with two columns:
`enroll_template_id, verify_template_id`.

`*.matches` (verification baseline) and `*.candidate_lists` (identification
baseline) preserve the upstream output schema — see the IJB-A documentation for
details.

## files.csv

A flat index of every image with columns `file_name, kind, extension, size_bytes,
referenced_in_protocols`. Use it for quick joins or coverage checks without
walking the filesystem.

## Local Stats

- Still images (`img/`): 5396
- Video frames (`frame/`): 20369
- Total images: 25765
- Distinct files referenced by protocols: 25791
- Verification splits: 10 (`IJB-A_11_sets/split1..split10`, 30 CSVs total)
- Verification baseline outputs: 10 (`IJB-A_11_output/`, 10 files total)
- Identification splits: 10 (`IJB-A_1N_sets/split1..split10`, 30 CSVs total)
- Identification baseline outputs: 10 (`IJB-A_1N_output/`, 10 files total)

## Missing protocol references

The upstream CleanData drop is missing 26 frames that are
referenced by at least one protocol CSV. Filter rows on these `FILE` values
before loading them, or skip silently:

- `frame/28264_00000.png`
- `frame/28296_00337.png`
- `frame/28296_00397.png`
- `frame/28296_00448.png`
- `frame/28296_00508.png`
- `frame/28296_00928.png`
- `frame/28296_00983.png`
- `frame/28296_01011.png`
- `frame/28296_01079.png`
- `frame/28296_01375.png`
- `frame/28296_01380.png`
- `frame/28296_01440.png`
- `frame/28296_01500.png`
- `frame/28296_01560.png`
- `frame/28317_01035.png`
- `frame/28332_01020.png`
- `frame/28496_00000.png`
- `frame/28552_00660.png`
- `frame/28593_00000.png`
- `frame/28789_00000.png`
- `frame/28928_00000.png`
- `frame/29192_00000.png`
- `frame/29387_00125.png`
- `frame/29563_00000.png`
- `frame/30369_00000.png`
- `frame/30387_00060.png`

## Unpacking video frames

Video frames are shipped as a single `frame.tar` archive (uploading 20k tiny
PNGs to the Hub triggered server-side commit failures). After downloading the
repo, extract it once so the protocol `FILE` paths resolve:

```bash
tar xf frame.tar     # creates frame/<video_id>_<frame_offset>.png
```

## Loading

```python
import pandas as pd
import tarfile
from huggingface_hub import snapshot_download
from pathlib import Path
from PIL import Image

root = Path(snapshot_download(repo_id="marcelohaps/ijb-a", repo_type="dataset"))

# Extract frame.tar in place if not already extracted
if not (root / "frame").exists():
    with tarfile.open(root / "frame.tar") as tf:
        tf.extractall(root)

# 1:1 verification, split 1
metadata = pd.read_csv(root / "protocols/IJB-A_11_sets/split1/verify_metadata_1.csv")
comparisons = pd.read_csv(
    root / "protocols/IJB-A_11_sets/split1/verify_comparisons_1.csv",
    header=None, names=["enroll_template_id", "verify_template_id"],
)

# Open the first face referenced by template_id 109
row = metadata[metadata["TEMPLATE_ID"] == 109].iloc[0]
image = Image.open(root / row["FILE"])
```

## Notes

IJB-A is distributed under the IARPA Janus benchmark license. Check the original
dataset terms before publishing or redistributing it. This package preserves the
upstream filenames (case included), bounding boxes, and protocol files verbatim.