metadata
dataset_info:
features:
- name: video
dtype: string
- name: row_id
dtype: int64
- name: action_score
dtype: float64
- name: peaks
list:
- name: build_up
dtype: float64
- name: caption
dtype: string
- name: drop_off
dtype: float64
- name: peak_end
dtype: float64
- name: peak_start
dtype: float64
splits:
- name: train
num_bytes: 192697
num_examples: 1028
download_size: 108343
dataset_size: 192697
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
⚡ FLASH
Fast Labelled Action Segment Highlights
Overview
FLASH is an open-source video evaluation dataset that tests a model's ability to identify the peak frames in a video given a user's prompt, for example, "The moment when the person lands on the ground". This is a much more challenging task than standard temporal action localization.
Analysis
We compare our dataset to the THUMOS'14 temporal action localization evaluation dataset. Our results are as follows:
+----------------------------------------------------+
| Statistics for Peak Duration |
+-----------+---------|--------+----------+-------+-------+-------+--------|
| dataset | count | mean | median | q25 | q75 | min | max |
|-----------+---------+--------+----------+-------+-------+-------+--------|
| FLASH | 1526 | 0.18 | 0.10 | 0.03 | 0.19 | 0.00 | 4.67 |
| THUMOS'14 | 3007 | 4.04 | 2.90 | 1.50 | 5.50 | 0.20 | 118.10 |
+-----------+---------+--------+----------+-------+-------+-------+--------+
Acknowledgements
This dataset is heavily built on the ActivityNet Captions dataset.