The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Bangla Sign Language (BdSL) Sentence-Level Dataset Collection Guidelines v2.0

Document Information

Version: 1.0
Last Updated: January 2026
Purpose: Standardized guidelines for collecting a linguistically valid, model-agnostic Bangla Sign Language sentence-level dataset
Scope: Sentence-level continuous sign language recognition and understanding

1. Introduction

1.1 Overview

This document establishes comprehensive standards for creating a Bangla Sign Language (BdSL) sentence-level dataset. The guidelines are designed to ensure:

Signal fidelity: High-quality visual and spatial data
Linguistic validity: Authentic BdSL grammar and structure
Model agnosticism: Usability across different recognition architectures
Long-term utility: Future-proof data representation

1.2 Target Applications

Sentence-level sign language recognition (SLR)
Sign language translation (SLT)
Linguistic analysis of BdSL
Multi-modal sign language understanding

1.3 Design Principles

Raw data preservation: Always retain original RGB video
Temporal integrity: Maintain natural signing dynamics
Linguistic authenticity: Respect BdSL grammar, not spoken Bengali
Signer independence: Prevent identity-based overfitting
Reproducibility: Enable verification and replication

2. Dataset Objectives

2.1 Primary Goals

Create a minimum x sentence corpus with x+ signers per sentence
Cover x+ unique BdSL glosses in natural sentential contexts
Support continuous recognition with co-articulation and transitions
Enable signer-independent model evaluation

2.2 Coverage Requirements

Category	Target	Notes
Total sentences	≥x	Unique sentence instances
Unique sentence types	≥x	Distinct grammatical patterns
Vocabulary size	x glosses	Canonical gloss inventory
Signers	≥20	Demographically diverse
Repetitions per sentence	5–10 signers	Cross-signer validation

2.3 Linguistic Diversity

Include sentences covering:

Sentence types: Declarative, interrogative (yes/no, WH-questions), imperative, negative
Temporal markers: Past, present, future
Spatial reference: Pointing, locative signs, directional verbs
Complexity levels: 3–15 glosses per sentence
Domain coverage: Daily life, education, health, social interaction

3. Technical Specifications

3.1 Video Capture Requirements

3.1.1 Camera Specifications

Parameter	Requirement	Rationale
Resolution	Minimum 1280×720 (720p) Recommended: 1920×1080 (1080p)	Ensures hand landmarks occupy 50–100+ pixels
Frame rate	Fixed 30 FPS or 60 FPS (No VFR)	Temporal consistency for sequence models
Codec	H.264/H.265 with high bitrate (≥8 Mbps for 1080p)	Minimizes compression artifacts
Color space	YUV 4:2:0 or higher	Standard for ML pipelines
Shutter speed	≥1/60s (to avoid motion blur)	Preserves hand shape clarity

3.1.2 Camera Positioning

Height: Chest to eye level of the signer (typically 1.2–1.5m from ground)
Distance: 1.5–2.5 meters from signer
Angle: Perpendicular to signer (0° horizontal deviation)
Stability: Tripod-mounted (no handheld recording)

3.1.3 Framing Guidelines

┌─────────────────────────────────┐
│    [20% margin]                 │
│  ┌─────────────────────┐        │
│  │   Head (full)       │        │
│  │   Upper torso       │        │
│  │   Hands (complete)  │        │
│  │   Signing space     │        │
│  └─────────────────────┘        │
│    [20% margin]                 │
└─────────────────────────────────┘

Vertical: Top of head to waist + 20% margin
Horizontal: Full signing space (typically shoulder-width × 2) + 20% margin
Depth of field: Ensure hands remain in focus throughout signing space

4. Recording Environment Standards

4.1 Physical Environment

4.1.1 Background

Type: Plain, matte, solid-color backdrop
Recommended colors: Light gray (RGB: 200,200,200), light blue (RGB: 180,200,220), or off-white
Prohibited:
- Textured walls or patterned surfaces
- Windows or reflective surfaces
- Moving objects or other people
- High-contrast backgrounds (pure white/black)

4.1.2 Lighting Setup

Standard Configuration:

Primary light: Diffused front lighting (softbox or umbrella)
Key-to-fill ratio: 2:1 to 3:1
Color temperature: 5000K–5600K (daylight balanced)
Avoid:
- Direct sunlight or window light
- Strong backlighting
- Harsh shadows on hands or face
- Flickering (use flicker-free LED panels)

Lighting Test:

Record 5-second test clip
Verify no harsh shadows on hands
Check for consistent brightness across signing space

4.2 Acoustic Environment

Record in a quiet environment (background noise <40 dB)
While audio is not the primary modality, clear audio helps with annotation alignment

5. Signing Performance Protocol

5.1 Pre-Recording Briefing

5.1.1 Signer Instructions

Provide signers with:

Written prompt in Bengali (for comprehension)
BdSL gloss sequence (for signing reference)
Video example from native BdSL user (if available)

Example:

Bengali prompt: "আমি স্কুলে যাই"
BdSL gloss: I যাই স্কুল
Instruction: Sign naturally in BdSL, not word-by-word Bengali

5.1.2 Practice Protocol

Allow 2–3 practice attempts before recording
Provide feedback on framing and naturalness
Do NOT provide feedback on signing style (preserve natural variation)

5.2 Recording Protocol

5.2.1 Temporal Structure

Each recording must follow this structure:

[Neutral pose] → [Transition] → [Sentence] → [Transition] → [Neutral pose]
  (0.5–1s)        (natural)     (content)     (natural)        (0.5–1s)

Neutral Pose Specification:

Hands resting naturally at sides or in lap
Relaxed facial expression
Facing camera directly

5.2.2 Signing Execution

DO:

✓ Sign at natural, conversational pace
✓ Use natural co-articulation and transitions
✓ Include appropriate non-manual features (facial expressions, head movements)
✓ Maintain hands within frame throughout
✓ Use natural BdSL grammar

DO NOT:

✗ Insert artificial pauses between signs
✗ Exaggerate or slow down signing unnaturally
✗ Follow spoken Bengali word order
✗ Use fingerspelling unless linguistically appropriate
✗ Look away from camera (except for gaze as linguistic feature)

5.2.3 Repetition Protocol

Record 3 takes minimum per sentence per signer
Select best take for inclusion (or keep all for data augmentation)
If errors occur, re-record the entire sentence (not partial segments)

5.3 Hand Visibility Requirements

5.3.1 Mandatory Conditions

Hands visible for ≥90% of frames during active signing
Both hands fully visible during two-handed signs
Hand size ≥50 pixels width for ≥80% of frames

5.3.2 Acceptable Occlusion

Temporary occlusion (<1 second) is acceptable if:

Linguistically motivated (e.g., signs articulated at face)
Hands remain trackable before and after occlusion

5.3.3 Re-recording Triggers

Re-record if:

Hand(s) move out of frame
Extended occlusion (>1s continuous)
Signer shifts position significantly

6. Linguistic and Annotation Standards

6.1 Gloss Vocabulary Design

6.1.1 Canonical Gloss List

Develop a standardized gloss inventory before data collection begins
One gloss per distinct BdSL sign (not per Bengali word)

Gloss Selection Principles:

Use the most common Bengali translation as gloss label
If multiple Bengali words → one sign, choose one canonical gloss
Use descriptive labels for classifier constructions: CL:PERSON-WALK, CL:VEHICLE-MOVE

6.1.2 Compound Signs

For lexicalized compounds, use hyphenation:

MOTHER-FATHER (parents)
EAT-FINISH (completed eating)

6.1.3 Handling Ambiguity

If a sign has multiple meanings:

Use context-neutral gloss: WORK (not JOB vs LABOR)
Document polysemy in gloss dictionary

6.2 Annotation Format

6.2.1 Primary Annotation

Each video requires:

{
  "video_id": "BdSL_S012_0347",
  "signer_id": "S012",
  "sentence_id": "0347",
  "gloss_sequence": ["আমি", "যাই", "স্কুল"],
  "bengali_text": "আমি স্কুলে যাই",
  "english_translation": "I go to school",
  "sentence_type": "declarative",
  "duration_seconds": 3.2,
  "fps": 30,
  "resolution": "1920x1080"
}

6.2.2 Non-Manual Feature (NMF) Annotation

Annotate sentence-level NMFs:

{
  "nmf_tags": {
    "question": false,
    "negation": false,
    "emphasis": false,
    "conditional": false,
    "topic_marker": false
  },
  "facial_expression": "neutral",
  "head_movement": "none",
  "body_lean": "none"
}

6.2.3 Temporal Boundaries (Optional Enhanced Annotation)

For research datasets, optionally include:

{
  "temporal_segments": [
    {"gloss": "NEUTRAL", "start_frame": 0, "end_frame": 15},
    {"gloss": "I", "start_frame": 16, "end_frame": 32},
    {"gloss": "SCHOOL", "start_frame": 33, "end_frame": 58},
    {"gloss": "GO", "start_frame": 59, "end_frame": 85},
    {"gloss": "NEUTRAL", "start_frame": 86, "end_frame": 96}
  ]
}

Note: Frame-level annotation is time-intensive; prioritize gloss sequence accuracy.

6.3 Annotation Quality Control

6.3.1 Annotator Requirements

Native BdSL users or certified interpreters
Training on gloss inventory and annotation tool
Inter-annotator agreement ≥85% (Cohen's kappa ≥0.80)

6.3.2 Validation Process

Initial annotation: Primary annotator
Review: Secondary annotator checks 20% of dataset
Reconciliation: Resolve disagreements via consensus or expert review
Final verification: Spot-check by dataset lead

7. Participant Diversity Requirements

7.1 Signer Demographics

7.1.1 Minimum Diversity Requirements

Dimension	Target Distribution	Minimum Count
Total unique signers	20–50	20
Gender	40–60% each	8 each
Age groups	18–30, 31–50, 50+	5 per group
Handedness	80–90% right, 10–20% left	3 left-handed
Native BdSL users	≥70%	14
BdSL fluency	Fluent/Native level	All
Geographic origin	2+ regions of Bangladesh	2 regions

7.1.2 Signer Background Documentation

Collect metadata for each signer:

{
  "signer_id": "S012",
  "age_group": "18-30",
  "gender": "female",
  "dominant_hand": "right",
  "bdsl_acquisition": "native",
  "years_signing": 18,
  "region": "Dhaka",
  "education_level": "undergraduate",
  "consent_obtained": true,
  "date_recorded": "2026-01-15"
}

7.2 Dataset Splitting Strategy

7.2.1 Signer-Independent Splits

Mandatory: No signer overlap between splits

Training set:   60% of signers (12 signers)
Validation set: 20% of signers (4 signers)
Test set:       20% of signers (4 signers)

7.2.2 Sentence Distribution

Ensure each split contains:

Representative samples of all sentence types
Balanced gloss frequency distribution
Similar sentence length distributions

7.2.3 Split Validation

After splitting, verify:

No signer leakage between sets
Vocabulary coverage >90% in each split
Sentence type balance (Chi-square test, p>0.05)

8. Quality Assurance Framework

8.1 Automated Quality Checks

8.1.1 Technical Quality Metrics

Run automated scripts to validate:

Check	Threshold	Action if Failed
Frame rate consistency	<2% frame drops	Reject video
Resolution compliance	Exactly as specified	Reject video
Hand detection rate	>90% frames	Flag for review
Hand size adequacy	>80% frames ≥50px	Flag for review
Landmark stability (jitter)	<10px/frame (90th percentile)	Flag for review
Video duration	2–30 seconds	Flag for review
File corruption	Zero errors	Reject video

8.1.2 Landmark Quality Assessment

Using MediaPipe or similar:

Extract hand landmarks for all frames
Compute confidence scores
Calculate frame-to-frame displacement
Flag videos with:
- Mean confidence <0.6
- 10% missing detections
- Sudden jumps >100px between frames

8.2 Manual Quality Review

8.2.1 Reviewer Checklist

For each video, verify:

Signer follows neutral-start-end protocol
Signing appears natural and fluent
No unnecessary pauses or repetitions
Hands remain in frame
Background and lighting meet standards
Gloss annotation matches signed content
Gloss sequence follows temporal order
Bengali and English translations are accurate
NMF tags are appropriate

8.2.2 Error Categories and Actions

Error Type	Severity	Action
Annotation mismatch	High	Re-annotate
Poor video quality	High	Re-record
Incomplete sentence	High	Re-record
Minor framing issue	Medium	Accept with flag
Slight lighting variation	Low	Accept

8.3 Inter-Annotator Agreement

Measure agreement on 100 randomly sampled videos
Compute Cohen's kappa for gloss sequence
Target: κ ≥ 0.80
If κ < 0.80, provide additional annotator training

9. Data Storage and Organization

9.1 File Naming Convention

BdSL_[SignerID]_[SentenceID]_[TakeNumber].mp4

Examples:
BdSL_S012_0347_T01.mp4
BdSL_S012_0347_T02.mp4
BdSL_S025_0128_T01.mp4

9.2 Directory Structure

BdSL_Dataset_v2.0/
│
├── videos/
│   ├── raw/                     # Original recordings (preserve forever)
│   │   ├── S001/
│   │   │   ├── BdSL_S001_0001_T01.mp4
│   │   │   ├── BdSL_S001_0001_T02.mp4
│   │   │   └── ...
│   │   ├── S002/
│   │   └── ...
│   │
│   └── processed/               # Cropped/normalized (optional)
│       └── ...
│
├── annotations/
│   ├── glosses/
│   │   ├── train.json
│   │   ├── val.json
│   │   └── test.json
│   │
│   ├── metadata/
│   │   └── sentence_metadata.json
│   │
│   └── temporal/                # Optional frame-level annotations
│       └── ...
│
├── features/                    # Derived features (not primary data)
│   ├── mediapipe_landmarks/
│   │   └── ...
│   └── other_features/
│       └── ...
│
├── documentation/
│   ├── gloss_dictionary.json   # Canonical gloss inventory
│   ├── signer_metadata.json    # Demographic info (anonymized)
│   ├── collection_log.csv      # Recording sessions
│   └── README.md
│
└── splits/
    ├── train_signers.txt
    ├── val_signers.txt
    └── test_signers.txt

9.3 Annotation File Format

9.3.1 Gloss Annotation JSON Schema

{
  "dataset_version": "2.0",
  "split": "train",
  "total_samples": 1200,
  "annotation_date": "2026-01-15",
  "annotators": ["A001", "A002"],
  
  "samples": [
    {
      "video_id": "BdSL_S012_0347_T01",
      "video_path": "videos/raw/S012/BdSL_S012_0347_T01.mp4",
      "signer_id": "S012",
      "sentence_id": "0347",
      "take_number": 1,
      
      "gloss_sequence": ["I", "SCHOOL", "GO"],
      "gloss_count": 3,
      
      "text_bengali": "আমি স্কুলে যাই",
      "text_english": "I go to school",
      
      "sentence_type": "declarative",
      "complexity_level": "simple",
      
      "nmf": {
        "question": false,
        "negation": false,
        "emphasis": false
      },
      
      "video_metadata": {
        "duration_seconds": 3.2,
        "fps": 30,
        "resolution": "1920x1080",
        "total_frames": 96
      },
      
      "quality_flags": {
        "manual_review": false,
        "automated_checks_passed": true,
        "landmark_quality": 0.92
      }
    }
  ]
}

9.4 Backup and Versioning

Maintain at least 3 copies of raw video data (local + 2 cloud)
Use version control for annotations (Git LFS or DVC)
Document all dataset modifications in changelog
Assign DOI for public release (e.g., via Zenodo)

10. Evaluation and Validation Protocol

10.1 Baseline Model Requirements

Before releasing the dataset, establish baseline performance:

10.1.1 Minimum Baseline Models

Train at least two baseline models:

Sequence-to-sequence model: (e.g., CTC-based recognizer)
- Input: MediaPipe landmarks or raw frames
- Output: Gloss sequence
Transformer-based model: (e.g., Transformer encoder-decoder)
- Input: Visual features
- Output: Gloss sequence

10.1.2 Evaluation Metrics

Report the following metrics on test set:

Metric	Definition	Target
Word Error Rate (WER)	`(S + D + I) / N`	<50% for baseline
Substitution Rate	Incorrect gloss predictions	Report
Deletion Rate	Missing glosses	Report
Insertion Rate	Extra glosses	Report
Sequence Accuracy	Exact match percentage	Report
Mean Levenshtein Distance	Average edit distance	Report

Where:

S = Substitutions
D = Deletions
I = Insertions
N = Total reference glosses

10.2 Cross-Validation Strategy

For signer-independent evaluation:

For each test fold:
  - Train on 12 signers (60%)
  - Validate on 4 signers (20%)
  - Test on 4 signers (20%)
  
Report mean and standard deviation across folds

10.3 Error Analysis

Conduct systematic error analysis:

Confusion matrix for frequent glosses
Error rate by sentence length
Error rate by sentence type
Error rate by signer demographics

10.4 Dataset Statistics

Document and report:

Statistic	Value
Total videos	[count]
Total unique sentences	[count]
Total unique glosses	[count]
Mean sentence length (glosses)	[value] ± [std]
Mean video duration (seconds)	[value] ± [std]
Gloss frequency distribution	[plot/table]
Sentence type distribution	[plot/table]

11. Ethical Considerations

11.1 Informed Consent

11.1.1 Consent Requirements

All signers must provide written informed consent covering:

Purpose of data collection
How videos will be used (research, model training, public release)
Data retention and sharing policies
Right to withdraw participation
Compensation (if applicable)
Privacy and anonymization measures

11.1.2 Consent Form Language

Provide consent forms in:

Bengali (written)
BdSL (video explanation with interpreter)

11.2 Privacy and Anonymization

11.2.1 Personally Identifiable Information (PII)

Videos show faces (required for NMF) but no names or identifiable metadata
Assign anonymous signer IDs (S001, S002, etc.)
Remove EXIF/metadata from video files
Store signer demographic data separately with restricted access

11.2.2 Data Access Control

Raw videos: Research team only
Processed landmarks: Shareable with collaborators
Public release: Only with explicit consent for public use

11.3 Community Engagement

11.3.1 Deaf Community Involvement

Consult with Bangladesh Deaf community organizations
Employ Deaf researchers/annotators where possible
Share results and benefits with community
Acknowledge community contributions

11.3.2 Cultural Sensitivity

Respect BdSL as a complete language (not "deficient" Bengali)
Avoid pathologizing or deficit-based framing
Use identity-first language ("Deaf person" not "hearing impaired")

11.4 Data Licensing

Recommended license:

CC BY-NC-SA 4.0 (Creative Commons Attribution-NonCommercial-ShareAlike)
Or Open Data Commons Open Database License (ODbL)

Clearly specify:

Attribution requirements
Commercial use restrictions (if applicable)
Derivative work conditions

12. Appendices

Appendix A: Equipment Checklist

Minimum Equipment

Camera: 1080p, 30fps minimum (DSLR, mirrorless, or high-end webcam)
Tripod: Stable, adjustable height
Lighting: 2× LED panels or softboxes (5000K–5600K)
Backdrop: 2m × 3m solid-color fabric or roll
Computer: Video preview and recording
Storage: ≥500GB per 1000 videos (raw)

Recommended Equipment

Camera: 1080p or 4K, 60fps (Sony A7, Canon R series, Panasonic GH5)
Lens: 35–50mm equivalent, f/2.8 or better
Lighting: 3-point lighting setup with diffusers
Audio recorder: Lavalier mic (for annotation reference)
Color checker: X-Rite ColorChecker for calibration

Appendix B: Gloss Dictionary Template

{
  "gloss": "SCHOOL",
  "bengali_translations": ["স্কুল", "বিদ্যালয়"],
  "english_translation": "school",
  "sign_type": "lexical",
  "handshape": "flat-hand",
  "movement": "tap-twice",
  "location": "non-dominant-palm",
  "description": "Dominant flat hand taps twice on non-dominant palm",
  "video_example": "glosses/SCHOOL_example.mp4",
  "frequency": 127,
  "related_glosses": ["COLLEGE", "UNIVERSITY"]
}

Appendix C: Sample Recording Session Plan

Session Duration: 2 hours
Signers per session: 2
Sentences per signer: 50

Time	Activity	Duration
0:00	Setup and equipment check	15 min
0:15	Signer briefing and consent	15 min
0:30	Practice recordings (5 sentences)	10 min
0:40	Recording Block 1 (25 sentences)	30 min
1:10	Break	10 min
1:20	Recording Block 2 (25 sentences)	30 min
1:50	Review and backup	10 min

Appendix D: Quality Assurance Checklist

Pre-Recording

Camera settings verified (resolution, fps, codec)
Lighting tested (no harsh shadows)
Background clean and appropriate
Signer positioned correctly
Test recording reviewed

During Recording

Monitor framing continuously
Check landmark detection (if real-time available)
Ensure neutral start/end for each clip
Mark problematic takes for re-recording

Post-Recording

Verify all files saved correctly
Run automated quality checks
Backup raw videos immediately
Update recording log

Annotation Phase

Gloss sequences verified by second annotator
NMF tags reviewed
Translations checked for accuracy
Metadata complete

Appendix E: Troubleshooting Guide

Issue	Possible Cause	Solution
Hand detection fails	Poor lighting, occlusion	Adjust lighting; re-record
Landmark jitter	Low resolution, motion blur	Increase resolution; adjust shutter speed
Inconsistent frame rate	Camera settings	Set fixed fps; avoid auto modes
Background clutter	Textured wall	Use plain backdrop
Signer out of frame	Incorrect positioning	Reframe; use viewfinder markings

Appendix F: Bibliography and Resources

Sign Language Dataset References

Jiang, S., Sun, B., Wang, L., et al. (2021). "Skeleton Aware Multi-modal Sign Language Recognition." CVPR.
Joze, H. R. V., Koller, O. (2019). "MS-ASL: A Large-Scale Data Set and Benchmark for Understanding American Sign Language." BMVC.
Duarte, A., et al. (2021). "How2Sign: A Large-scale Multimodal Dataset for Continuous American Sign Language." CVPR.

BdSL Linguistic Resources

Zeshan, U. (2003). "Indo-Pakistani Sign Language Grammar: A Typological Outline."
M. A. (2024)."IsharaKhobor: A Large-Scale Sentence-Level Bangla Sign Language Dataset."

Technical Guidelines

MediaPipe Documentation: https://google.github.io/mediapipe/
Sign Language Processing Resources: https://sign-language-processing.github.io/

Document Revision History

Version	Date	Changes	Author
1.0	2026-01	Initial guidelines	Team

End of Document

Downloads last month: 2

Total file size:

28 kB