---
pretty_name: Enhanced Fall Dataset
language:
  - en
task_categories:
  - video-classification
tags:
  - fall-detection
  - accident-detection
  - video
size_categories:
  - 10K<n<100K
---

# Enhanced Fall Dataset (total = 25,691)

## Prerequisite

Download ```bear7011/gemma-4-e4b-kinetics_54K``` first for some overlapped videos. (This setting prevents cascading forgetting.)

## Notification!

1. Do not mix the sora-accident dataset into the training process.
2. Video sources:
    - `videos/kinetics_fall`, `videos/kinetics_neg` — [Kinetics dataset](https://github.com/cvdfoundation/kinetics-dataset)
    - `videos/oops` — [OOPS! dataset (Columbia)](https://oops.cs.columbia.edu/data/)

## File Structure

```
├── annotations
│   ├── prompts.json
│   ├── train.json
│   └── val.json
├── distribution.json
├── label.txt
├── README.md
└── videos
    ├── kinetics_fall
    ├── kinetics_neg
    └── oops
```

## Label Distribution

This dataset is composed of three types of labels:
1. General Fall (8,000)
2. Hard Negative, similar to fall (10,691)
    - Normal Hard Negative (3,800)
    - Hard Hard Negative (3,500)
    - Extreme Hard Negative (3,391)
3. Overlapped videos from 54k (7,000)