OmarNabill/event-detection

Event Detection: video classification dataset

A small, labeled video dataset for classifying a whole video clip into one event class. It is built by merging several public datasets into one folder with unified file names, a single format (.mp4), and a fixed train/val/test split.

Work in progress. This is version 1 and it contains only 3 of the 6 planned classes. Running, Crowd Formation and Fire/Smoke videos will be added in future updates. Statistics, the split and the file list will change when new classes are added.

Classes

Class	Status
Fighting	included
Fainting/Falling	included
Other	included
Running	coming soon
Crowd Formation	coming soon
Fire/Smoke	coming soon

Other means "not similar to the target events" (normal activity, no need to identify what it is).

Roadmap

• Add Running videos.
• Add Crowd Formation videos.
• Add Fire/Smoke videos.
• Add more Fighting sources (all Fighting videos currently come from a single source).
• Add more Other videos covering diverse everyday scenes.
• Rebuild the train/val/test split after the new classes are added, possibly with a stricter group-based split.

Structure

videos/
  fainting_falling/   <source>_<id>.mp4
  fighting/
  other/
metadata.csv

metadata.csv columns:

Column	Description
file_name	relative path of the video
label	Fighting / Fainting/Falling / Other
source	original dataset (gmdcsa24, fight_surv, le2i)
group	source or scene name (use it for stricter group-based splits)
frames, fps, duration_s	video info
width, height	resolution
size_mb	file size
orig_name	original file name
split	train / val / test

Statistics of the current version (587 videos, about 1.2 GB)

Label	Source	Videos
Fainting/Falling	gmdcsa24	79
Fainting/Falling	le2i	97
Fighting	fight_surv	150
Other	fight_surv	150
Other	gmdcsa24	81
Other	le2i	30

Split: 410 train / 88 val / 89 test (about 70/15/15, stratified by label and source, seed 42).

Usage

from datasets import load_dataset

ds = load_dataset("OmarNabill/event-detection", split="train")   # loads all rows
train = ds.filter(lambda x: x["split"] == "train")
val   = ds.filter(lambda x: x["split"] == "val")
test  = ds.filter(lambda x: x["split"] == "test")

Or read metadata.csv directly with pandas and open the videos by file_name.

Sources

• GMDCSA24: https://github.com/ekramalam/GMDCSA24-A-Dataset-for-Human-Fall-Detection-in-Videos (Fall -> Fainting/Falling, ADL -> Other)
• fight-detection-surv-dataset: https://github.com/seymanurakti/fight-detection-surv-dataset (fight -> Fighting, noFight -> Other)
• Le2i / ImViA Fall Detection Dataset: https://imvia.ube.fr/?p=346 (used through a Kaggle mirror: https://www.kaggle.com/datasets/tuyenldvn/falldataset-imvia). Only Coffee_room_01/02 and Home_01/02 are included. A video is labeled Fainting/Falling when its official annotation file marks a fall window, otherwise Other. Office and Lecture_room videos were left out because the mirror has no annotation files for them.

Processing

• Videos with a missing or unreadable file were dropped, and exact duplicates were removed.
• All videos were re-encoded to .mp4 (H.264, no audio) where needed (Le2i .avi); original resolution and fps are kept.
• Labels are whole-video labels. Le2i fall videos also contain frames before and after the fall.

Limitations

• Small dataset (hundreds of videos), the test split has only 89 videos, so results are noisy.
• The split is random at the video level, not at the person or scene level, so the same scene or actor can appear in train and test. Scores may be optimistic.
• All Fighting videos come from a single source.
• Resolution and fps vary a lot between sources (320x180 up to 1920x1080, 10 to 30 fps), so resize and sample frames in your dataloader.
• Falls in Le2i and GMDCSA24 are mostly staged.

License and terms

Each source dataset has its own license and terms of use. This repository does not grant any license on top of them. Check the original licenses (for example, Le2i is listed under CC-BY-NC-SA 3.0 in other benchmark papers, so verify it yourself) before sharing this repo publicly or using it commercially. Please cite the original datasets when you use this data.