kirandevraj/ISL-Fingerspelling

ISL Fingerspelling Dataset

First continuous fingerspelling dataset for Indian Sign Language (ISL).

Paper: "Continuous Fingerspelling Dataset for Indian Sign Language" (WSLP @ AACL-IJCNLP 2025)

Statistics

Segments	Duration	Characters
1,308	70.85 min	14,685

Character count is taken from letter_annotations, where every character instance is individually annotated.

Structure

├── videos/                        # 1308 mp4 files
├── fingerspelling_annotations.csv # Segment annotations (word level)
├── fingerspelling_annotations_updated.csv # Same, transcribed from the letter annotations
├── letter_annotations.json        # Character-level frame boundaries
├── letter_annotations.csv         # Same, as a flat table
├── letter_timestamps.csv          # Letter boundaries with source-video timestamps
├── detection_eval_segments.csv    # Ground truth for the detection evaluation
├── detection_eval_other_fs.csv    # Additional fingerspelling inside the eval windows
├── automatic_segments_detection.csv      # Discovered segments, detection pipeline
├── automatic_segments_sliding_window.csv # Discovered segments, sliding-window pipeline
├── split_info.csv                 # Standard train/test split
├── signer_split.csv               # Signer-independent train/test split
└── localization_timestamps.csv    # Temporal localization in source videos

fingerspelling_annotations.csv

Maps video segments to their transcriptions.

Column	Description
uid	Unique segment ID ({video_id}_seg{index}_signer)
text	Fingerspelled text

fingerspelling_annotations_updated.csv

The segment transcription read directly off the character-level annotations: for each segment, the letters concatenated in order. Same schema as fingerspelling_annotations.csv (uid, text), so it can be used in its place.

It differs from the original transcription in 164 of the 1,308 segments. Because it is derived from the frame-by-frame annotation, it reflects what is actually fingerspelled in the video — including cases where a signer spells only part of a word, or omits the space between two words.

letter_annotations.json / letter_annotations.csv

Character-level frame boundaries for every fingerspelled letter, introduced in the ICPR 2026 paper below. For each segment, every character instance is marked with the frame at which it starts and the frame at which it ends.

Boundaries follow a transition-onset rule: after a letter is formed the hand holds that shape briefly, then begins moving toward the next letter. The boundary is placed at the frame where that motion begins, so a single frame ends one character and starts the next.

Statistic	Value
Segments	1,308
Character instances	14,685
Annotated frames	125,476
Unique characters	38
Avg. frames per character	8.5

The JSON is keyed by uid (the video is videos/{uid}.mp4):

{
  "-CpDPjXeB2c_seg000": [
    {"letter": "v", "start_frame": 0, "end_frame": 10},
    {"letter": "i", "start_frame": 10, "end_frame": 19}
  ]
}

The CSV carries the same content one row per character instance:

Column	Description
uid	Segment ID, matching fingerspelling_annotations.csv
char_index	Position of the character within the segment
letter	The character (may be a space or punctuation)
start_frame	First frame of the character
end_frame	Frame at which the next character begins

Quality: validated by an independent professional ISL interpreter on 660 balanced letter samples. 92.0% agreement on letters shown in isolation; after reviewing disagreements in context, validated accuracy is 99.1% (6 genuine errors in 660).

Relationship to fingerspelling_annotations.csv

These annotations were produced by reviewing every segment frame by frame, and in the process a number of spelling errors in the original word-level transcripts were corrected. Where the two files disagree, letter_annotations is the authoritative transcription — it reflects what is actually fingerspelled in the video.

signer_split.csv

The signer-independent partition introduced in the ICPR paper: test signers do not appear in training. 810 train / 498 test over three signers. Use this instead of split_info.csv to measure generalisation to unseen signers.

Column	Description
uid	Segment ID
signer	Signer identifier (signer_00 … signer_02)
split	train or test

letter_timestamps.csv

The same character boundaries expressed as timestamps in the original source video, rather than as frame indices within the extracted segment.

Column	Description
video_id	YouTube video ID
segment_index	Segment index within the video
uid	Segment ID
signer	Signer identifier
signer_split	train / test for the signer-independent split
standard_split	train / test for the standard split
transcript	Fingerspelled text of the segment
letter	The character
abs_start_sec, abs_end_sec	Character boundaries in the source video
start_frame, end_frame	Character boundaries within the segment
rel_start_sec, rel_end_sec	Character boundaries relative to the segment
fps	Frame rate of the source video

Segment-level boundaries are not repeated here; they are in localization_timestamps.csv.

This file also carries signer_split, the signer-independent partition introduced in the ICPR paper, in which test signers do not appear in training.

detection_eval_segments.csv

Ground truth for the fingerspelling detection evaluation reported in the ICPR paper (precision / recall / F1 and downstream CER). 204 fingerspelling segments across 92 source videos, each marked with its boundaries in the original video.

Column	Description
video_id	YouTube video ID
uid	Segment ID, matching fingerspelling_annotations.csv
word	Fingerspelled word
start_sec, end_sec, duration	Segment boundaries in the source video

Evaluation uses a 10-second window centred on each of these segments. Frames inside an annotated fingerspelling span are positives and the rest of the window is negative.

detection_eval_other_fs.csv

Within those 10-second windows, the target segment is often not the only fingerspelling present. This file records 203 additional fingerspelling regions, annotated manually across 73 of the 92 evaluation videos: acronyms, initials, numbers and short forms such as TN, 2Y, 600, 25O2019.

These must be treated as positives too. Without them, a detector that correctly fires on one of these regions is penalised as a false positive.

Column	Description
video_id	YouTube video ID
start_sec, end_sec, duration_sec	Region boundaries in the source video
frames, fps	Frame count and frame rate of the region
transcript	What is fingerspelled
letter_count	Number of characters

Some of these are very short. In the example shown in the paper, an M lasting 0.16s falls below the 0.5s minimum-duration filter and is missed by every method, which is what pulls recall down.

automatic_segments_detection.csv

Fingerspelling segments discovered automatically in the iSign corpus (127,105 videos with sentence-level transcripts but no sign-level annotation). The frame classifier detects candidate regions, these are transcribed, and the transcription is matched against proper nouns extracted from the sentence transcript.

The video files for these segments are not distributed here. uid refers to a clip in the iSign dataset, and start_sec / end_sec locate the segment within that clip. Obtain the videos from iSign: Joshi et al., iSign: A Benchmark for Indian Sign Language Processing (arXiv:2407.05404).

7,479 segments, of which 4,736 are exact matches. These have no character-level frame annotations — they carry a segment-level word label only, and are intended as additional weakly-labelled training data.

Column	Description
uid	Discovered segment ID
source_id	iSign source video ID
matched_word	Word matched from the sentence transcript
transcription	Model transcription of the region
cer	Character error rate between the two (%)
all_matches	All candidate matches considered, with their CER
start_sec, end_sec, duration	Region boundaries in the source video
mean_confidence	Mean frame-classifier confidence over the region
transcript_text	Sentence transcript the word was drawn from

Quality: in a blind manual transcription study of 100 randomly sampled segments, 79% were exact matches (mean CER 7.42%).

automatic_segments_sliding_window.csv

The second discovery pipeline over the same iSign corpus. Multi-scale windows are slid across each video, transcribed, and matched against candidate words from the sentence transcript; matches are accepted below 30% CER. 7,833 segments.

Like the detection-based segments, these carry a word-level label only and have no character-level frame annotations. As above, the underlying videos come from the iSign dataset and are not redistributed here.

Column	Description
uid	iSign clip ID
word	Word matched from the sentence transcript
start_sec, end_sec, duration	Segment boundaries within the clip

Quality: in a blind manual transcription study of 100 randomly sampled segments, 71% were exact matches (mean CER 7.98%).

Combined with the detection-based pipeline this gives over 11,700 segments, a 9x increase over the manually annotated set. The two pipelines overlap: 3,867 of the detection-based segments (51.7%) are not found by the sliding window.

localization_timestamps.csv

Contains temporal boundaries of fingerspelling segments in the original YouTube videos.

Column	Description
video_id	YouTube video ID
segment_index	Segment index within the video
start_time	Start timestamp (HH:MM:SS.mmm)
start_sec	Start time in seconds
end_time	End timestamp (HH:MM:SS.mmm)
end_sec	End time in seconds
duration_str	Duration (HH:MM:SS.mmm)
duration_sec	Duration in seconds
transcript	Fingerspelled text

Citation

The character-level frame annotations:

@inproceedings{kirandevraj2026densefs,
  title={Dense Frame Annotations for Low-Resource ISL Fingerspelling Recognition},
  author={Kirandevraj, R and Kurmi, Vinod K and Namboodiri, Vinay P and Jawahar, CV},
  booktitle={ICPR},
  year={2026}
}

The dataset:

@inproceedings{kirandevraj2025islfingerspelling,
  title={Continuous Fingerspelling Dataset for Indian Sign Language},
  author={Kirandevraj, R and Kurmi, Vinod K and Namboodiri, Vinay P and Jawahar, CV},
  booktitle={WSLP @ AACL-IJCNLP},
  year={2025}
}

Contact

kirandevraj.r@research.iiit.ac.in

Project Page

https://kirandevraj.github.io/ISL-Fingerspelling