Note

This is a community plugin, an external project maintained by its respective author. Community plugins are not part of FiftyOne core and may change independently. Please review each plugin’s documentation and license before use.

GitHub Repo

Video to Frames Plugin for FiftyOne#

A FiftyOne plugin that converts a video dataset into a frames view: one image sample per video frame, with all frame-level labels, at the frame rate and size you choose.

The frames view lives on the same dataset as your videos. Label edits made on frames are synced back to the videos, so you don’t need a separate image dataset.

Installation#

fiftyone plugins download https://github.com/mgustineli/video-to-frames-plugin

FiftyOne Enterprise#

Plugins are installed on the deployment by an admin:

  1. Download the repo as a zip: video-to-frames-plugin-main.zip

  2. In the App, go to Settings → Plugins → Install plugin and upload the zip

Or with the SDK, as an admin:

import fiftyone.management as fom

fom.upload_plugin("video-to-frames-plugin-main.zip", overwrite=True)

Background (delegated) runs need a delegated operator orchestrator on the deployment. Without one, choose Execute now when running the operator.

Usage#

  1. Open a video dataset in the App

  2. Click Convert video to frames in the samples grid actions (or find it in the operator browser)

  3. Pick the options and run:

Option

Description

Target

The whole dataset, the current view, or the selected samples

Frame rate

All frames, Fixed FPS (exactly N frames per second) or Max FPS (only downsample videos above N)

Max frame height

Resize larger frames, keeping the aspect ratio (e.g. 4K → 1080p)

Output directory

Where to write the frame images, e.g. gs://my-bucket/frames. By default, a folder next to each video

Re-sample existing frames

Frames that were already sampled are reused; enable this to re-sample them (e.g. at a new size)

Saved view name

The frames view is saved on the dataset under this name

When run immediately, the App opens the frames view when it’s done. When run in the background (delegated), open the saved view from the view selector once the run finishes.

Frames are sampled once. The saved view only references the sampled images, so loading it later is fast.

Notes#

  • Sampling decodes every video with ffmpeg, so long or high-resolution videos take a while. Background (delegated) execution is the default.

  • The output location must be writable by the deployment (or by your SDK credentials when running locally). With cloud media, frames are written back to the same bucket by default.

  • If a video can’t be sampled, the operation fails with an error instead of creating a view with missing images.

  • Frames that already exist are not resized. To change the size of frames you already sampled, enable Re-sample existing frames.

SDK equivalent#

The operator wraps to_frames():

import fiftyone as fo

dataset = fo.load_dataset("my-videos")

# Sample frames at 1 fps, at most 1080p tall
dataset.to_frames(sample_frames=True, fps=1, max_size=(-1, 1080))

# Save a frames view that uses the sampled images
frames = dataset.to_frames(fps=1)
dataset.save_view("frames-1fps-1080p", frames)

License#

MIT