README.md

metadata

language:
  - en
tags:
  - sign language recognition
  - emergency response
  - computer vision

CLARIS - Critical Emergency Sign Language Dataset

This dataset is a curated subset of the "Google - Isolated Sign Language Recognition" dataset, specifically filtered for the CLARIS (Clear and Live Automated Response for Inclusive Safety) project.

Dataset Description

The primary goal of the CLARIS project is to develop a mobile application that provides a lifeline for the Deaf community during emergencies. This dataset was created to train a proof-of-concept AI model capable of recognizing a vocabulary of critical emergency-related signs.

The data consists of pre-extracted landmark coordinates from video clips of isolated signs. It originates from the Google - Isolated Sign Language Recognition Kaggle Competition.

Dataset Structure

The dataset is provided in both CSV and Parquet (coming soon) formats. Each row represents the coordinates of a single landmark in a single frame of a video sequence.

Column	Dtype	Description
frame	int16	The frame number within the sequence.
row_id	object	A unique identifier for the landmark within the frame.
type	object	The type of landmark (face, left_hand, right_hand, pose).
landmark_index	int16	The index of the landmark within its type.
x	float64	The normalized x-coordinate of the landmark.
y	float64	The normalized y-coordinate of the landmark.
z	float64	The normalized z-coordinate of the landmark (depth).
path	object	The path to the original source parquet file for the sequence.
participant_id	int64	A unique identifier for the participant (signer).
sequence_id	int64	A unique identifier for the sign sequence.
sign	object	The ground truth label for the sign being performed.

Curation Process

To create a focused dataset for our specific use case, we performed a two-step curation process:

1. Vocabulary Filtering: We selected 62 signs deemed most relevant for describing medical, fire, or intruder emergencies.
2. Participant Filtering: To create a manageable dataset for rapid prototyping, we constrained the data to sequences from two distinct participants who had a balanced distribution of the target signs.

This process resulted in a final dataset containing 1,719 unique sign sequences, comprising over 37 million landmark rows.

Usage

We recommend using the Parquet file for faster loading times.

import pandas as pd

# Load the full curated dataset
df = pd.read_parquet('claris_curated_dataset.parquet')

# Or load the smaller, subsampled version
df_sample = pd.read_parquet('claris_subsample_dataset.parquet')

print(df.head())

Link to Project Notebook

The complete methodology, including data preprocessing, model training, and analysis, can be found in our Kaggle notebook: https://www.kaggle.com/code/eveelyn/datathon2025-med