File size: 6,440 Bytes
947ee22
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-nc-sa-4.0
task_categories:
- video-classification
- image-classification
language:
- fr
tags:
- sign_language
- sign_language_processing
- sign_language_recognition
- isolated_sign_language_recognition
- computer_vision
- natural_language_processing
pretty_name: LSFB-ISOL (Isolated French Belgian Sign Language Dataset)
configs:
  - config_name: "500"
    data_files:
      - split: train
        path:
          - "shards/500/shard_000003.tar"
          - "shards/500/shard_000004.tar"
          - "shards/500/shard_000005.tar"
          - "shards/500/shard_000006.tar"
          - "shards/500/shard_000007.tar"
          - "shards/500/shard_000008.tar"
          - "shards/500/shard_000009.tar"
      - split: validation
        path:
          - "shards/500/shard_000001.tar"
          - "shards/500/shard_000002.tar"
      - split: test
        path: "shards/500/shard_000000.tar"
  - config_name: "750"
    data_files:
      - split: train
        path:
          - "shards/750/shard_000003.tar"
          - "shards/750/shard_000004.tar"
          - "shards/750/shard_000005.tar"
          - "shards/750/shard_000006.tar"
          - "shards/750/shard_000007.tar"
          - "shards/750/shard_000008.tar"
          - "shards/750/shard_000009.tar"
      - split: validation
        path:
          - "shards/750/shard_000001.tar"
          - "shards/750/shard_000002.tar"
      - split: test
        path: "shards/750/shard_000000.tar"
  - config_name: "2000"
    data_files:
      - split: train
        path:
          - "shards/2000/shard_000003.tar"
          - "shards/2000/shard_000004.tar"
          - "shards/2000/shard_000005.tar"
          - "shards/2000/shard_000006.tar"
          - "shards/2000/shard_000007.tar"
          - "shards/2000/shard_000008.tar"
          - "shards/2000/shard_000009.tar"
      - split: validation
        path:
          - "shards/2000/shard_000001.tar"
          - "shards/2000/shard_000002.tar"
      - split: test
        path: "shards/2000/shard_000000.tar"
  - config_name: all
    data_files:
      - split: train
        path:
          - "shards/all/shard_000003.tar"
          - "shards/all/shard_000004.tar"
          - "shards/all/shard_000005.tar"
          - "shards/all/shard_000006.tar"
          - "shards/all/shard_000007.tar"
          - "shards/all/shard_000008.tar"
          - "shards/all/shard_000009.tar"
      - split: validation
        path:
          - "shards/all/shard_000001.tar"
          - "shards/all/shard_000002.tar"
      - split: test
        path: "shards/all/shard_000000.tar"
---

# LSFB-ISOL: Isolated French Belgian Sign Language Dataset

LSFB-ISOL is the **isolated-sign** subset of the [LSFB corpus](https://www.corpus-lsfb.be/): one pre-segmented, single-sign clip per sample, extracted from the continuous ([lsfb-cont](https://huggingface.co/datasets/ppoitier/lsfb-cont)) recordings. It is built for **Isolated Sign Language Recognition (ISLR)**.

<video controls src="https://huggingface.co/datasets/ppoitier/lsfb-isol/resolve/main/example.mp4" width="360"></video>

## Dataset structure

The dataset is distributed as [WebDataset](https://github.com/webdataset/webdataset) `.tar` shards, split into four vocabulary-size variants that share the same underlying data:

| Variant | Vocabulary |
|---|---|
| `500`   | 500 most frequent signs |
| `750`   | 750 most frequent signs |
| `2000`  | 2000 most frequent signs |
| `all`   | Full vocabulary |

Each variant is split into `train` (shards 3-9), `validation` (shards 1-2) and `test` (shard 0).

Each sample is keyed by a clip id (e.g. `CLSFBI0301A_S008_B_12100_12437`) and contains:

- `pose.<body_part>.npy` — keypoint sequence for one body part (`upper_pose`, `left_hand`, `right_hand`, `lips`, `left_eye`, `right_eye`, `left_eyebrow`, `right_eyebrow`, `left_iris`, `right_iris`)
- `label.txt` / `label.idx` — the sign gloss (e.g. `bonjour`) and its class index
- `signer.txt` — anonymized signer id (e.g. `S008`)
- `language.txt` — always `lsfb`
- `boundaries.json``start_ms` / `end_ms` / `start_frame` / `end_frame` of the clip within its parent recording
- `linguistic_metadata.json``gloss`, `lemma`, `sign_type`
- `parent.txt` — id of the source continuous recording (see [lsfb-cont](https://huggingface.co/datasets/ppoitier/lsfb-cont))

Raw video clips are stored separately in `videos.tar`, indexed by `videos.tar.index.json` (byte offset/length per clip), so a single clip can be extracted without downloading the whole archive. `example.mp4` is a preview clip.

## Usage

This dataset is designed to be loaded with [sign-language-data-loading](https://github.com/ppoitier/sign-language-data-loading) (`sldl`, [PyPI](https://pypi.org/project/sign-language-data-loading/)), which reads the WebDataset shards directly and handles pose/video decoding, windowing and target encoding:

```bash
pip install sign-language-data-loading
```

```python
from sldl import SignLanguageDataset
from sldl.configs import LSFBIsolConfig

dataset = SignLanguageDataset.from_config(
    LSFBIsolConfig(
        root="path/to/lsfb-isol",  # local clone or download of this repo
        variant="500",
        split="training",
    )
)

sample = dataset[0]
print(sample["label"])                             # e.g. "bonjour"
print(sample["poses"]["upper_pose"].shape)
```

Pose/annotation transforms (e.g. converting sign boundaries to frame labels) are provided by [sign-language-tools](https://github.com/ppoitier/sign-language-tools) (`slt`, [PyPI](https://pypi.org/project/sign-language-tools/)), which `sldl`'s target encoders build on.

## License

This dataset is distributed under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).

## Related projects

- [sign-language-data-loading](https://github.com/ppoitier/sign-language-data-loading) ([PyPI](https://pypi.org/project/sign-language-data-loading/)) — PyTorch-friendly data loading for sign language datasets stored as WebDataset shards; this dataset's intended loader.
- [sign-language-tools](https://github.com/ppoitier/sign-language-tools) ([PyPI](https://pypi.org/project/sign-language-tools/)) — pose and annotation transforms for sign language processing.
- [lsfb-cont](https://huggingface.co/datasets/ppoitier/lsfb-cont) — the continuous (untrimmed) counterpart of this dataset, useful for Continuous Sign Language Recognition (CSLR) and Sign Language Segmentation (SLS).