Overcoming Small Data Limitations in Video-Based Infant Respiration Estimation

arXiv ID: 2512.06888v1

Published: 2025-12-07

Authors: Liyang Song, Hardik Bishnoi, Sai Kumar Reddy Manne, Sarah Ostadabbas, Briana J. Taylor, Michael Wan

Categories: cs.CV

Relevance Score: 0.95 / 1.00

View on arXiv Download PDF

Summary

This paper addresses the critical lack of resources for developing contactless infant respiration monitoring by introducing the AIR-400 dataset, the largest public video dataset for infant respiration estimation. It also develops the first reproducible pipelines utilizing infant-specific ROI detection and spatiotemporal neural processing with optical flow, establishing crucial state-of-the-art benchmarks to overcome small data limitations.

Medical Relevance

Contactless infant respiration monitoring can enable early detection of breathing irregularities, which are critical indicators for conditions like neurodevelopmental impairments and Sudden Infant Death Syndrome (SIDS), improving outcomes and potentially saving lives.

AI Health Application

The AI application is a contactless, vision-based system for real-time infant respiration estimation. This system aims to use computer vision algorithms and neural processing on video data to monitor an infant's breathing, enabling early detection of irregularities that could lead to medical intervention and improve health outcomes.

Key Points

Identifies a significant gap: robust computer vision algorithms and datasets for infant respiration estimation are lacking, unlike for adults.
Introduces AIR-400, an Annotated Infant Respiration dataset of 400 videos, expanding the public corpus by 275 new, carefully annotated videos from 10 subjects.
Develops the first reproducible pipelines for infant respiration estimation, featuring infant-specific region-of-interest (ROI) detection.
The pipelines utilize spatiotemporal neural processing, further enhanced by optical flow as an input to improve accuracy.
Establishes the first reproducible benchmarks for vision-based infant respiration estimation through comprehensive experimental validation.
The authors make their dataset, code repository, and trained models publicly available to foster further research and development.
Aims to enable advances in early detection and treatment of infant breathing irregularities, linked to neurodevelopmental impairments and SIDS.

Methodology

The study involved creating a new dataset (AIR-400) by recruiting 10 subjects and carefully annotating 275 new videos, combined with existing data. They developed novel reproducible pipelines for infant respiration estimation based on infant-specific region-of-interest detection. These pipelines employ spatiotemporal neural processing, incorporating optical flow as an input feature to enhance model performance. Comprehensive experiments were conducted to evaluate and establish benchmarks.

Key Findings

The paper successfully introduced AIR-400, a significantly expanded public dataset for infant respiration. It developed the first effective and reproducible vision-based infant respiration estimation pipelines. Through rigorous experimentation, it established the first state-of-the-art benchmarks, demonstrating the feasibility and potential of contactless monitoring despite small data limitations.

Clinical Impact

This research provides foundational tools (dataset, algorithms, benchmarks) for developing reliable contactless respiration monitors for infants. This could lead to earlier diagnosis and intervention for breathing irregularities, potentially reducing the incidence of SIDS and mitigating neurodevelopmental impairments, offering a safer and less intrusive monitoring solution in clinical and home settings.

Limitations

The abstract itself doesn't explicitly state limitations of *their own* work, but rather frames the problem by highlighting prior limitations in the field: a scarcity of large public video datasets with annotated respiration data for infants and a lack of effective, reproducible algorithms.

Future Directions

The public release of the dataset, code, and models strongly implies a future direction of encouraging the research community to build upon these foundations, develop more robust algorithms, and expand the dataset further to improve the accuracy and generalizability of infant respiration estimation systems for real-world clinical applications.

Medical Domains

pediatrics neonatology respiratory medicine developmental neurology sleep medicine

Keywords

infant respiration monitoring contactless sensing computer vision small data video dataset deep learning optical flow SIDS

Abstract

The development of contactless respiration monitoring for infants could enable advances in the early detection and treatment of breathing irregularities, which are associated with neurodevelopmental impairments and conditions like sudden infant death syndrome (SIDS). But while respiration estimation for adults is supported by a robust ecosystem of computer vision algorithms and video datasets, only one small public video dataset with annotated respiration data for infant subjects exists, and there are no reproducible algorithms which are effective for infants. We introduce the annotated infant respiration dataset of 400 videos (AIR-400), contributing 275 new, carefully annotated videos from 10 recruited subjects to the public corpus. We develop the first reproducible pipelines for infant respiration estimation, based on infant-specific region-of-interest detection and spatiotemporal neural processing enhanced by optical flow inputs. We establish, through comprehensive experiments, the first reproducible benchmarks for the state-of-the-art in vision-based infant respiration estimation. We make our dataset, code repository, and trained models available for public use.