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

This repository contains the annotation and split definitions for OmniFall, a comprehensive benchmark that unifies eight public indoor fall datasets under a consistent ten-class annotation scheme, complemented by the OOPS-Fall benchmark of genuine accidents captured in the wild.

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

Also have a look for additional information on our project page:

OmniFall Project Page

Overview

Falls are the leading cause of fatal injuries among older adults worldwide. While the mechanical event of falling lasts only a fraction of a second, the critical health risk often comes from the ensuing "fallen" state—when a person remains on the ground, potentially injured and unable to call for help.

OmniFall addresses three critical limitations in current fall detection research:

1. Unified Taxonomy: Rather than binary fall/no-fall classification, we provide a consistent ten-class scheme across datasets that distinguishes transient actions (fall, sit down, lie down, stand up) from their static outcomes (fallen, sitting, lying, standing).

2. Combined Benchmark: We unify eight public datasets (14+ hours of video, 112 subjects, 31 camera views) into a single benchmark with standardized train/val/test splits.

3. In-the-Wild Evaluation: We include OOPS-Fall, curated from genuine accident videos of the OOPS dataset to test generalization to real-world conditions.

Datasets

This benchmark includes annotations for the following datasets:

1. CMDFall (7h 25m single view) - 50 subjects, 7 synchronized views
2. UP Fall (4h 35m) - 17 subjects, 2 synchronized views
3. Le2i (47m) - 9 subjects, 6 different rooms
4. GMDCSA24 (21m) - 4 subjects, 3 rooms
5. CAUCAFall (16m) - 10 subjects, 1 room
6. EDF (13m) - 5 subjects, 2 views synchronized
7. OCCU (14m) - 5 subjects, 2 views not synchronized
8. MCFD (12m) - 1 subject, 8 views
9. OOPS-Fall - Curated subset of genuine fall accidents from the OOPS dataset, strong variation in subjects and views.

Structure

The repository is organized as follows:

• labels/ - CSV files containing frame-level annotations for each dataset as well as label2id.csv
• splits/ - Train/validation/test splits for cross-subject (CS) and cross-view (CV) evaluation
  • splits/cs/ - Cross-subject splits, where training, validation, and test sets contain different subjects
  • splits/cv/ - Cross-view splits, where training, validation, and test sets contain different camera views

Label Format

Each label file in the labels/ directory follows this format:

path,label,start,end,subject,cam,dataset
path/to/clip,class_id,start_time,end_time,subject_id,camera_id,dataset_name

Where:

• path: Relative path to the video, given the respective dataset root.
• label: Class ID (0-9) corresponding to one of the ten activity classes:
  • 0: walk
  • 1: fall
  • 2: fallen
  • 3: sit_down
  • 4: sitting
  • 5: lie_down
  • 6: lying
  • 7: stand_up
  • 8: standing
  • 9: other
• start: Start time of the segment (in seconds)
• end: End time of the segment (in seconds)
• subject: Subject ID
• cam: Camera view ID
• dataset: Name of the dataset

For OOPS-Fall, only fall segments and non-fall segments are labeled; non-falls are labels as "other", independent of the underlying content, as long as it is not a fall. Cam and subject ids in OOPS-Fall are -1.

Split Format

Split files in the splits/ directory list the video segments included in each partition. You can use the split paths to filter the label data.:

path
path/to/clip

Evaluation Protocols

We provide multiple evaluation configurations via the dataset.yaml file:

Basic Configurations

• default: Access to all dataset labels (huggingface loads everything into the train split by default.)
• cs: Cross-subject splits for all datasets
• cv: Cross-view splits for all datasets

Individual Dataset Configurations

• caucafall, cmdfall, edf, gmdcsa24, le2i, mcfd, occu, up_fall, OOPS: Access to individual datasets with their respective cross-subject splits

Multi-Dataset Evaluation Protocols

• cs-staged: Cross-subject splits combined across all staged datasets
• cv-staged: Cross-view splits combined across all staged datasets
• cs-staged-wild: Train and validate on staged datasets with cross-subject splits, test on OOPS-Fall
• cv-staged-wild: Train and validate on staged datasets with cross-view splits, test on OOPS-Fall

Examples

from datasets import load_dataset
import pandas as pd

# Load the datasets
print("Loading datasets...")

# Note: We separate segment labels and split definitions, but hugginface datasets always expects a split.
# Thats why all labels are in the train split when loaded, but we create the actual splits afterwards.
labels = load_dataset("simplexsigil2/omnifall", "labels")["train"]

cv_split = load_dataset("simplexsigil2/omnifall", "cv")
cs_split = load_dataset("simplexsigil2/omnifall", "cs")

# There are many more splits, relevant for the paper:
# - cv-staged -> Only lab datasets
# - cs-staged -> Only lab datasets
# - cv-staged-wild -> Lab datasets for train and val, only OOPS-Fall in test set
# - cs-staged-wild -> Lab datasets for train and val, only OOPS-Fall in test set

# Convert to pandas DataFrames
labels_df = pd.DataFrame(labels)
print(f"Labels dataframe shape: {labels_df.shape}")

# Process each split type (CV and CS)
for split_name, split_data in [("CV", cv_split), ("CS", cs_split)]:
    print(f"\n{split_name} Split Processing:")

    # Process each split (train, validation, test)
    for subset_name, subset in split_data.items():
        # Convert to DataFrame
        subset_df = pd.DataFrame(subset)

        # Join with labels on 'path'
        merged_df = pd.merge(subset_df, labels_df, on="path", how="left")

        # Print statistics
        print(f"  {subset_name} split: {len(subset_df)} videos, {merged_df.dropna().shape[0]} labelled segments")

        # Print examples
        if not merged_df.empty:
            print(f"\n  {subset_name.upper()} EXAMPLES:")
            random_samples = merged_df.sample(min(3, len(merged_df)))
            for i, (_, row) in enumerate(random_samples.iterrows()):
                print(f"  Example {i+1}:")
                print(f"    Path: {row['path']}")
                print(f"    Start: {row['start']}")
                print(f"    End: {row['end']}")
                print(f"    Label: {row['label']}")
                print(f"    Subject: {row['subject']}")
                print(f"    Dataset: {row['dataset']}")
                print(f"    Camera: {row['cam']}")
                print()

Label definitions

In this section we provide additional information about the labelling process to provide as much transparency as possible.

There are 10 labels:

• 0|walk - Move around, including jogging and running and "drunk walking", but only if it is not part of some special exercise like pulling your knees up. Not when pushing a large object like a chair, but included carrying something small like an apple.
• 1|fall - The act of falling (from any previous state). Includes falling on a bed, if the process is not a controlled lying down with arms as support.
• 2|fallen - Being on the ground or a mattress after a fall.
• 3|sit_down - Sitting down on bed or chair or ground.
• 4|sitting - Sitting on bed or chair or ground.
• 5|lie_down - Lying down intentionally (in contrast to a fall) on ground or bed.
• 6|lying - Being in a lying position (in bed or on the ground) after intentionally getting into that position.
• 7|stand_up - Standing up from a fallen state, from lying or sitting. Includes getting from lying position into sitting position.
• 8|standing - Standing around without walking.
• 9|other - Any other activity, including e.g. walking while pushing an object like a chair.

Motion Types

There are two types of motions, dynamic ones like walk or stand_up and static ones like fallen, sitting, lying.

Generally we annotate dynamic motions as soon as the first frame appears which belongs to that action.

Let's say we see a person walk, then fall. The first frame which indicates a motion which does seem to be different to walk is the start of fall. Sometimes it is necessary to have a look at a person walking to learn when the motion begins to change to something else.

For static motions the label begins with the first frame where the person comes to a resting state. For sit_down, the label ends when the person reaches a state where it is no longer adjusting its body position but comes to a rest. fall ends when the person is no longer moving caused by the inertia of the fall. The following fallen might contain movement on the ground, but no movement which belongs to fall or stand_up.

Label Sequences

There are some natural sequences of labels like fall, fallen and stand_up. However, it is not always the case that these appear together. Sometimes the person might directly stand up again without any time at rest on the ground, in this case there is no fallen segment. Likewise sometimes there is no sitting segment.

Lying down can be on a bed or on the ground, it is intentional in contrast to fall. There are falls which follow sit_down or lie_down if it is from a chair or from a bed.

When a person is lying in a bed and getting up to sit in the bed we label this as stand_up, even if the person is still sitting in the bed. A sequence could then be lying, stand_up, sitting, stand_up to describe a person which first lies down, then gets into a sitting position, waits a little, then gets fully up.

Sometimes it is not perfectly clear if it is sit_down followed by lie_down or simply lie_down. This depends on there being a moment of rest or not. If the person spends a short amount of time in the sitting position it is the former, if the person directly goes from sit_down to lie_down without rest it is only labeled lie_down. Similar thoughts apply to stand_up.

In the video at the bottom of this page we show how we leveraged VGG VIA to perform the annotations. Note, that we pre-load the original dataset labels as visual aid, but relabel the whole video with our label definitions. In the video below, CMDFall is shown, which already provides relatively detailed labels, the original labels of other datasets are more sparse. Additionally, even CMDFall does not label all frames, but only specific segments, while our labels cover mostly every frame.

The blurred out regions were added by us in post-processing to protect the subjects privacy on this page, they are not part of the original videos.

Citation

If you use OmniFall in your research, please cite our paper (will be updated soon) as well as all sub-dataset papers:

@misc{omnifall,
      title={OmniFall: A Unified Staged-to-Wild Benchmark for Human Fall Detection}, 
      author={David Schneider and Zdravko Marinov and Rafael Baur and Zeyun Zhong and Rodi Düger and Rainer Stiefelhagen},
      year={2025},
      eprint={2505.19889},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.19889}, 
},

@inproceedings{omnifall_cmdfall,
  title={A multi-modal multi-view dataset for human fall analysis and preliminary investigation on modality},
  author={Tran, Thanh-Hai and Le, Thi-Lan and Pham, Dinh-Tan and Hoang, Van-Nam and Khong, Van-Minh and Tran, Quoc-Toan and Nguyen, Thai-Son and Pham, Cuong},
  booktitle={2018 24th International Conference on Pattern Recognition (ICPR)},
  pages={1947--1952},
  year={2018},
  organization={IEEE}
},

@article{omnifall_up-fall,
  title={UP-fall detection dataset: A multimodal approach},
  author={Mart{\'\i}nez-Villase{\~n}or, Lourdes and Ponce, Hiram and Brieva, Jorge and Moya-Albor, Ernesto and N{\'u}{\~n}ez-Mart{\'\i}nez, Jos{\'e} and Pe{\~n}afort-Asturiano, Carlos},
  journal={Sensors},
  volume={19},
  number={9},
  pages={1988},
  year={2019},
  publisher={MDPI}
},

@article{omnifall_le2i,
  title={Optimized spatio-temporal descriptors for real-time fall detection: comparison of support vector machine and Adaboost-based classification},
  author={Charfi, Imen and Miteran, Johel and Dubois, Julien and Atri, Mohamed and Tourki, Rached},
  journal={Journal of Electronic Imaging},
  volume={22},
  number={4},
  pages={041106--041106},
  year={2013},
  publisher={Society of Photo-Optical Instrumentation Engineers}
},

@article{omnifall_gmdcsa,
  title={GMDCSA-24: A dataset for human fall detection in videos},
  author={Alam, Ekram and Sufian, Abu and Dutta, Paramartha and Leo, Marco and Hameed, Ibrahim A},
  journal={Data in Brief},
  volume={57},
  pages={110892},
  year={2024},
  publisher={Elsevier}
},

@article{omnifall_cauca,
  title={Dataset CAUCAFall},
  author={Eraso, Jose Camilo and Mu{\~n}oz, Elena and Mu{\~n}oz, Mariela and Pinto, Jesus},
  journal={Mendeley Data},
  volume={4},
  year={2022}
},

@inproceedings{omnifall_edf_occu,
  title={Evaluating depth-based computer vision methods for fall detection under occlusions},
  author={Zhang, Zhong and Conly, Christopher and Athitsos, Vassilis},
  booktitle={International symposium on visual computing},
  pages={196--207},
  year={2014},
  organization={Springer}
},

@article{omnifall_mcfd,
  title={Multiple cameras fall dataset},
  author={Auvinet, Edouard and Rougier, Caroline and Meunier, Jean and St-Arnaud, Alain and Rousseau, Jacqueline},
  journal={DIRO-Universit{\'e} de Montr{\'e}al, Tech. Rep},
  volume={1350},
  pages={24},
  year={2010}
},

@inproceedings{omnifall_oops,
  title={Oops! predicting unintentional action in video},
  author={Epstein, Dave and Chen, Boyuan and Vondrick, Carl},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  pages={919--929},
  year={2020}
}

License

The annotations and split definitions in this repository are released under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

The original video data belongs to their respective owners and should be obtained from the original sources.

Contact

For questions about the dataset, please contact [[email protected]].

How we used VGG VIA to densely annotate videos for Omnifall

We provide this video to shortly demonstrate how the annotation process was conducted, increasing transparency. Note, that the CMDFall dataset already provides realtively detailed labels, this is not the case for many of the other datasets.