OmniFall: A Unified Staged-to-Wild Benchmark for Human Fall Detection

2025-05-26

AI Summary

Existing fall detection research is hindered by small-scale, lab-controlled datasets exhibiting substantial domain bias, leaving real-world generalization capability largely uncharacterized. To address this, we introduce the first staged-to-wild unified benchmark for fall detection, integrating eight public datasets (56 hours of video) and establishing a ten-class fine-grained annotation schema with standardized evaluation protocols—including the novel OOPS-Fall subset comprising real-world accident footage. We propose a cross-dataset-compatible video segmentation annotation paradigm and an out-of-distribution generalization evaluation framework. Our empirical analysis reveals, for the first time, that state-of-the-art pretrained models—including I3D and VideoMAE—suffer over 40% accuracy degradation on real-world data. We fully open-source all data, annotations, and code, providing a reproducible, comparable, and robust evaluation foundation for future fall detection research.

Technology Category

• Computer Vision: Segmentation
• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• Security and Privacy: Data transparency and provenance
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Current video-based fall detection research mostly relies on small, staged datasets with significant domain biases concerning background, lighting, and camera setup resulting in unknown real-world performance. We introduce OmniFall, unifying eight public fall detection datasets (roughly 14 h of recordings, roughly 42 h of multiview data, 101 subjects, 29 camera views) under a consistent ten-class taxonomy with standardized evaluation protocols. Our benchmark provides complete video segmentation labels and enables fair cross-dataset comparison previously impossible with incompatible annotation schemes. For real-world evaluation we curate OOPS-Fall from genuine accident videos and establish a staged-to-wild protocol measuring generalization from controlled to uncontrolled environments. Experiments with frozen pre-trained backbones such as I3D or VideoMAE reveal significant performance gaps between in-distribution and in-the-wild scenarios, highlighting critical challenges in developing robust fall detection systems. OmniFall Dataset at https://huggingface.co/datasets/simplexsigil2/omnifall , Code at https://github.com/simplexsigil/omnifall-experiments

Problem

• Addresses domain biases in staged fall detection datasets
• Unifies diverse datasets for fair cross-dataset comparison
• Measures generalization from controlled to real-world scenarios

Innovation

• Unifies eight datasets under consistent taxonomy
• Provides complete video segmentation labels
• Establishes staged-to-wild evaluation protocol