GufranSabri/isharah-selfi

Selfi: Arabic Sign Language Dataset (Frames)

Selfi is a video-frame dataset for isolated/continuous Arabic Sign Language recognition, collected from multiple signers and organized into two evaluation protocols:

• SI (Signer-Independent) — dev/test signers do not appear in train.
• US (Unseen Sentences) — dev/test sentences (gloss sequences) do not appear in train, but signers may overlap with train.

The gloss vocabulary (960 glosses) is shared between SI and US.

Repository layout

Group1/                # raw extracted frames, signers 00 01 02 03 08 12 19
Group2/                # raw extracted frames, signers 08 10 11 18 20 21
Group3/                # raw extracted frames, signers 05 13 14 15 18 19
SI/
  train.txt
  dev.txt
  test.txt
  gloss_dict.npy
  train_info.npy
  dev_info.npy
  test_info.npy
US/
  train.txt
  dev.txt
  test.txt
  gloss_dict.npy
  train_info.npy
  dev_info.npy
  test_info.npy

Signer IDs are not unique across groups — e.g. signer 08 appears in both Group1 and Group2, and signer 18/19 appear in both Group2 and Group3. A sample is therefore only uniquely identified by its Group/signer/sample_id path, not by signer+sample_id alone.

Frame folders

GroupX/<signer>/<sample_id>/<frame_idx>.png

Example: Group2/08/0714/0143.png

• <signer>: 2-digit zero-padded signer ID (e.g. 08).
• <sample_id>: 4-digit zero-padded sample/sentence ID (e.g. 0714). This matches the numeric part of the id field in train.txt/dev.txt/test.txt (see below).
• <frame_idx>.png: 4-digit zero-padded frame index (e.g. 0143.png), one PNG per extracted video frame, in temporal order. Frame counts vary per sample (roughly 90–450 frames observed).

To load all frames of a sample in order:

import glob
frames = sorted(glob.glob(f"{dataset_root}/Group2/08/0714/*.png"))

Annotation files (SI/, US/)

train.txt, dev.txt, test.txt

Pipe-delimited, one header line then one row per sample:

id|gloss
08_0714|يوجد حراره وسط
08_0717|يوجد الم معده
...

• id: <signer>_<sample_id> (e.g. 08_0714). Note this ID alone does not tell you which Group the frames live in — use *_info.npy (below) to resolve the full frame path, since the same signer ID can exist in more than one group.
• gloss: ground-truth gloss sequence for the sample, space-separated, in signing order. This is the annotation used for training/evaluating CSLR models (order matters, no per-frame alignment is provided).

Row counts:

Split	SI	US
train	5,996	8,355
dev	1,499	329
test	1,499	310

gloss_dict.npy

A pickled Python dict saved with numpy.save(..., allow_pickle=True), shared between SI and US (960 glosses total). Maps each gloss string to a [index, frequency] pair:

import numpy as np
gloss_dict = np.load("SI/gloss_dict.npy", allow_pickle=True).item()

gloss_dict["يوجد"]        # -> [index, frequency]
# index: 1-based integer class ID used as the model's target label for this gloss.
# frequency: kept for compatibility with the original preprocessing script;
#            not populated with real counts for this release (currently 0 for every gloss).

Use the index value (first element) as the integer class label when mapping a gloss string to a training target; index 0 is reserved for the CTC blank token.

train_info.npy, dev_info.npy, test_info.npy

A pickled Python dict per split, saved with numpy.save(..., allow_pickle=True). This is the file that resolves an id from train.txt/dev.txt/test.txt to its actual frame folder. Keys are integer indices 0 .. N (where N = number of samples in the split); the last integer key is a placeholder/sentinel and should be ignored — consumers should iterate range(N) where N = len(info) - 1.

import numpy as np
info = np.load("SI/train_info.npy", allow_pickle=True).item()

info[0]
# {
#   'fileid': 'Group2/08/0714',
#   'folder': 'Group2/08/714/*.png',
#   'signer': '08',
#   'label': 'يوجد حراره وسط',
#   'sentence': 'لدي حراره متوسطه',
#   'original_info': 'Group2/08/0714|0|لدي حراره متوسطه|يوجد حراره وسط|يوجد حراره وسط|n/a\n'
# }

Field meanings:

• fileid: Group/signer/sample_id, zero-padded sample ID (matches the frame folder name).
• folder: a glob pattern for the frame files, relative to the dataset root (the frames/ directory containing Group1/Group2/Group3). ⚠️ The sample-ID path segment in this pattern is not zero-padded (e.g. Group2/08/714/*.png) and will not match any files as-is — you must zero-pad that segment to 4 digits before globbing:

  parts = info[i]["folder"].split("/")
  parts[-2] = parts[-2].zfill(4)          # '714' -> '0714'
  frame_glob = "/".join(parts)             # 'Group2/08/0714/*.png'
  frames = sorted(glob.glob(f"{dataset_root}/{frame_glob}"))
  

• signer: signer ID string (matches the <signer> path segment).
• label: gloss sequence (identical to the gloss column in the corresponding .txt file).
• sentence: the natural-language Arabic sentence the gloss sequence corresponds to (not used for CSLR training, provided for reference/translation tasks).
• original_info: the raw source annotation row this entry was generated from, kept for traceability; the trailing n/a field is unused.

Sample counts per split (len(info) - 1, excluding the sentinel key) match the row counts in the corresponding .txt file above.

Loading example (PyTorch-style)

import glob
import numpy as np

dataset_root = "."  # points at the folder containing Group1/ Group2/ Group3/
split = "train"
protocol = "SI"      # or "US"

info = np.load(f"{protocol}/{split}_info.npy", allow_pickle=True).item()
gloss_dict = np.load(f"{protocol}/gloss_dict.npy", allow_pickle=True).item()
num_samples = len(info) - 1  # last integer key is a sentinel, not a real sample

for idx in range(num_samples):
    sample = info[idx]

    # resolve frame folder (zero-pad the sample-id segment)
    parts = sample["folder"].split("/")
    parts[-2] = parts[-2].zfill(4)
    frame_glob = "/".join(parts)
    frame_paths = sorted(glob.glob(f"{dataset_root}/{frame_glob}"))

    # resolve integer labels for the gloss sequence
    labels = [gloss_dict[g][0] for g in sample["label"].split(" ") if g in gloss_dict]

Notes / gotchas

• Signer IDs repeat across groups — always resolve paths via *_info.npy, never by reconstructing Group{signer}/{sample_id} manually from the id column alone.
• The folder field in *_info.npy needs the zero-pad fix above before globbing; this is a quirk of how the info files were generated, not a data error.
• The last integer key in each *_info.npy dict is a non-sample sentinel — exclude it (i.e. use len(info) - 1 as the sample count, as shown above).
• gloss_dict.npy is identical between SI/ and US/ (same 960-gloss vocabulary); the frequency field is not populated with real counts in this release.