README.md

metadata

annotations_creators:
  - found
language_creators:
  - found
language:
  - en
license:
  - mit
multilinguality:
  - monolingual
pretty_name: CLIP-Kinetics700
size_categories:
  - 100K<n<1M
task_categories:
  - feature-extraction
  - zero-shot-classification

Dataset Card for CLIP-Kinetics70

Table of Contents

• Table of Contents
• Dataset Description
  • Dataset Summary
  • Dataset Preprocessing
• Dataset Structure
  • Data Instances
  • Data Fields
  • Data Splits
• Dataset Creation
  • Source Data
• Simple Experiments
  • Zero-shot Evaluation
  • Linear-probe Evaluation

Dataset Description

Dataset Summary

CLIP-Kinetics700 is a compressed version of the Kinetics700 dataset using OpenAI's CLIP model.

The original dataset is ~700 GB making it difficult to use and hold in memory on one machine. By downsampling each video to 1 FPS and encoding the frames using CLIP we we're able to compress the dataset to ~8 GB making it very memory-friendly and easy to use.

Dataset Preprocessing

clip-video-encode is a tool you can use to easily and efficiently compute CLIP embeddings from video frames. We used it to generate the embeddings for this dataset.

Dataset Structure

Data Format

We formatted this as a WebDataset for better data-loading performance when training the models. Each split contains a list of tar files each with 10000 data samples. This format can be read and used easily using the EmbeddingWebDatasetReader from clip-video-encode.

CLIP-Kinetics700
 ├── splits.csv
 ├── ds_00000.tar
 |     ├── vid_00000.npy
 |     ├── vid_00000.txt
 |     ├── vid_00000.json
 |     ├── vid_00001.npy
 |     ├── vid_00001.txt
 |     ├── vid_00001.json
 |     └── ...
 |     ├── vid_10000.npy
 |     ├── vid_10000.txt
 |     ├── vid_10000.json
 ├── ds_00001.tar
 |     ├── vid_10001.npy
 |     ├── vid_10001.txt
 |     ├── vid_10001.json
 │     ...
 ...

Data Fields

• vid.npy: the numpy array with the per-frame embeddings. Shape -> (n_frames, 512)
• vid.cap: the "caption" of the video. In this case it is the Kinetics700 label.
• vid.json: additional metadata - YouTube video ID, start time, end time.

Data Splits

• Train - 536489 samples | 54 tar's
• Validation - 33966 samples | 4 tar's
• Test - 64532 samples | 7 tar's

Dataset Creation

Source Data

Data was sourced from DeepMind's Kinetics700 dataset and downloaded using this convenient repository.

Simple Experiments

Using this repository we evaluate CLIP-Kinetics700 with the following simple methods:

Zero-shot Evaluation

	Accuracy
Top-1	0.31
Top-5	0.56
mean(Top1, Top5)	0.44

Linear-probe Evaluation

	Accuracy
Top-1	0.41
Top-5	0.65
mean(Top1, Top5)	0.53