V-JEPA Mamba Dataset V2: Preprocessed Egocentric Video Frames for Pretraining

V-JEPA Mamba Dataset V2 contains preprocessed video frames from the Egocentric-10K dataset, formatted for V-JEPA Mamba pretraining. Each video is filtered to be at least 3 minutes long, clipped to the first 3 minutes, and processed to 1,440 frames at 8 fps. The dataset was created by author 'rookierufus' and was last updated on June 11, 2026.

Use Cases

• Pretraining video foundation models based on the described frame extraction and resolution.
• Self-supervised learning on egocentric video data based on the dataset's stated purpose.
• Benchmarking video representation learning methods using the standardized 384x384 resolution frames.

Strengths

• Videos are processed to a consistent 384x384 RGB resolution via shortest-side scaling and center cropping.
• Each processed video contains 1,440 frames, extracted at a rate of 8 frames per second.
• Videos are filtered for a minimum duration of 3 minutes, providing a baseline length for training sequences.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset's license is unspecified, which may restrict usage.

Provenance

Source

Egocentric-10K dataset

Collection Method

Videos were filtered, clipped to the first 3 minutes, and frames were extracted at 8 fps using ffmpeg.

Freshness

Last updated 2026-06-11 18:41:18; freshness should be verified.