---
license: cc-by-nc-4.0
task_categories:
- video-classification
- feature-extraction
language:
- en
tags:
- video
- streaming
- event-boundary-detection
- kinetics-gebd
size_categories:
- n<1K
dataset_info:
- config_name: default
  features:
  - name: video_id
    dtype: string
  - name: split
    dtype: string
  - name: source_split
    dtype: string
  - name: source_dataset
    dtype: string
  - name: category
    dtype: string
  - name: youtube_url
    dtype: string
  - name: clip_reference
    dtype: string
  - name: clip_start_seconds
    dtype: int64
  - name: clip_end_seconds
    dtype: int64
  - name: fps
    dtype: float64
  - name: duration_seconds
    dtype: float64
  - name: num_frames
    dtype: int64
  - name: annotator_count
    dtype: int64
  - name: selected_annotator_index
    dtype: int64
  - name: selected_annotator_consistency
    dtype: float64
  - name: mean_annotator_consistency
    dtype: float64
  - name: boundaries
    list:
    - name: start_seconds
      dtype: float64
    - name: end_seconds
      dtype: float64
    - name: boundary_kind
      dtype: string
    - name: change_type
      dtype: string
  - name: annotation_license
    dtype: string
  splits:
  - name: train
    num_examples: 6
  - name: validation
    num_examples: 2
  - name: test
    num_examples: 2
---

# Streaming GEBD Causal Preview

Ten metadata-only examples derived from the official Kinetics-GEBD raw
annotations. This preview validates the schema intended for causal streaming
event-boundary training before publishing a full processed release.

## Splits

- `train`: 6 examples from the original Kinetics-GEBD training annotations.
- `validation`: 2 examples from the original validation annotations.
- `test`: 2 held-out examples from the original validation annotations.

The selected annotation for each clip is the annotation from the annotator with
the highest consistency score. The complete release may replace this preview
policy with consensus boundary aggregation.

## Video access

This repository does not redistribute Kinetics YouTube videos. Each row
contains the YouTube ID, source URL, and the original Kinetics clip start/end
seconds. Availability and usage rights of each source video remain subject to
YouTube and the original uploader.

## Causal use

At time `t`, a streaming model must receive only frames with timestamps less
than or equal to `t`. It must not use frames after the candidate boundary.

`boundary_kind` separates:

- `semantic_event`: actor, object, action, or other semantic state change.
- `visual_change`: cut, pan, zoom, fade, or speed transition.

## Sources and license

Annotations are derived from
[StanLei52/GEBD](https://github.com/StanLei52/GEBD), licensed under
CC BY-NC 4.0. Cite:

```bibtex
@inproceedings{shou2021gebd,
  title={Generic Event Boundary Detection: A Benchmark for Event Segmentation},
  author={Shou, Mike Zheng and Lei, Stan Weixian and Wang, Weiyao and Ghadiyaram, Deepti and Feiszli, Matt},
  booktitle={ICCV},
  year={2021}
}
```