Datasets:
metadata
license: apache-2.0
task_categories:
- video-classification
- feature-extraction
tags:
- video
- tubelets
- egocentric
- raw-frames
- computer-vision
size_categories:
- 10K<n<100K
Tubelet Dataset
Pre-extracted raw video tubelets from the Egocentric-10K dataset
Tensor Format
Each file is a single tubelet saved with and can be loaded with .
| Property | Value |
|---|---|
| Shape | |
| Dtype | |
| Layout | — Frames × Height × Width × Channels |
| Frames | 64 raw RGB frames per tubelet |
| Resolution | 384 × 384 center-cropped |
| Pixel Range | 0–255 (unnormalized) |
1. Frame Rate Downsampling
- Source videos are downsampled to 4 FPS using FFmpeg's filter.
- The mode exactly matches Meta's frame selection.
2. Shorter-Side Scaling (Aspect-Ratio Preserving)
- The shorter side of each frame is scaled to 438 pixels while preserving the original aspect ratio.
- This value is derived from the canonical V-JEPA proportion: .
- For 224×224 crops, V-JEPA scales the shorter side to 256px. We scale proportionally for 384×384 crops.
3. Center Cropping
- A 384 × 384 center crop is taken from each scaled frame.
- FFmpeg's filter defaults to center cropping when no x/y offsets are specified.
4. Tubelet Chunking
- The full decoded frame sequence is divided into non-overlapping chunks of exactly 64 frames.
- Incomplete trailing chunks (fewer than 64 frames) are discarded.
FFmpeg Command Used
Directory Structure
Files are partitioned across multiple root directories (, , , ...) to comply with HuggingFace's 10,000 files per directory limit.
Usage
Source Dataset
- Source: builddotai/Egocentric-10K
- Content: Egocentric (first-person) video data from industrial/factory environments
- Factories: factory_001 through factory_050