titania7777 / VideoFrameSampler

Salient Video Frames Sampling Method Using the Mean of Deep Features for Efficient Model Training (KIBME 2021)

VideoFrameSampler

Salient Video Frames Sampler for Efficient Model Training Using the Mean of Deep Features

Summary

This code's purpose is to find meaningful frames in both trimmed and untrimmed video datasets. And this Sampler working only with UCF101, HMDB51, ActivityNet datasets.
We only provides video frame sampler codes(returns the JSON file), however, we will be published training codes which utilize this sampler results in another repository later!!.

Requirements

• opencv-python
• ffmpeg-python
• torch
• pillow-simd(optional)

Usage(UCF101)

Clone this repository

git clone https://github.com/titania7777/VideoFrameSampler.git

Download the dataset

cd ./VideoFrameSampler/Data/UCF101/
./download.sh

Run an Index Sampler

cd ../../
python sampler_run.py --dataset-name UCF101 --split-id 1

Loading Test

python sampler_test.py --dataset-name UCF101 --split-id 1 --sequence-length 16

Sampled Annotations

We provide our sampler results here

Download UCF101 Sampled Annotations

Download HMDB51 Sampled Annotations

Download ActivityNet Sampled Annotations

Examples on ActivityNet

Kayaking(Uniform)

Kayaking(Our)

Laying Tile(Uniform)

Laying Tile(Our)

Citation

If you use this code in your work, please cite our work

@inproceedings{SalientFrameSampler2021,
    author={Hyeok Yoon and Young-Gi Kim and Ji-Hyeong Han},
    title={Salient Video Frames Sampling Method Using the Mean of Deep Features for Efficient Model Training},
    booktitle={Proceedings of the 2021 Korean Institute of Broadcast and Media Engineers Summer Conference},
    pages={318-321},
    year={2021},
}