CLIP-Encoded Video Frames from Kinetics700 at 1 FPS

CLIP-Kinetics700 is a compressed version of the Kinetics700 dataset, using OpenAI's CLIP model to encode video frames. The original ~700 GB dataset was reduced to ~8 GB by downsampling videos to 1 frame per second. It was created by iejMac and last updated in July 2022.

Use Cases

• Perform zero-shot classification on CLIP-encoded video frames using text prompts.
• Use the compressed CLIP features for feature extraction in video understanding models.
• Analyze temporal patterns in video data using the 1 FPS frame sampling rate.

Strengths

• Significant size reduction from ~700 GB to ~8 GB, improving accessibility.
• Preprocessing with CLIP encoding provides ready-to-use features for machine learning.
• Based on the large-scale Kinetics700 dataset, which contains 700 human action classes.

Limitations

• The specific number of rows (videos or frames) is not provided.
• Downsampling to 1 FPS may lose high-frequency motion details critical for some action recognition tasks.
• The dataset's geographic and demographic bias is inherited from the original Kinetics700 source.

Provenance

Source

Compressed from the original Kinetics700 dataset.

Collection Method

Videos were downsampled to 1 frame per second and frames were encoded using OpenAI's CLIP model.

Freshness

Last updated on 2022-07-11.

Geography

Region tag suggests 'us', but specific coverage is unknown.