Adaptive video frame extraction for Structure from Motion, Gaussian Splatting, photogrammetry, and related 3D reconstruction workflows.
Frame Extractor turns video into image sequences for reconstruction pipelines such as COLMAP and Gaussian Splatting.
Unlike conventional fixed-interval extraction, it can select keyframes dynamically based on camera motion. This reduces redundant images when the camera is stationary or moving slowly, while extracting frames more frequently during fast movement or rotation.
The result is a smaller, more useful image set with better overlap between neighboring views.
A typical video-to-SfM workflow extracts every Nth frame:
Video → every 10th frame → images → SfM / Gaussian Splatting
The problem is that camera motion is rarely constant.
When the camera stops or moves slowly, fixed sampling creates many nearly identical images. When it moves or rotates quickly, the same interval may leave too much change between frames, making feature matching and pose recovery harder.
Frame Extractor adapts the spacing automatically:
Slow motion: ● ● ●
Fast motion: ● ● ● ● ● ● ● ●
- Motion-adaptive keyframe extraction with Low, Medium, and High presets
- Fixed-interval extraction when exact sampling is preferred
- Multiple timeline regions to extract only useful sections of a video
- Manual frame extraction from the preview
- JPEG or lossless PNG output at source resolution
- Frame-accurate navigation and live preview
- Optional separate output folders for each selected region
- CSV manifest, extraction settings, and summary saved alongside the images
- Desktop GUI and command-line interface
- All processing runs locally
Frame Extractor is written in C++20 and is designed to be fast and easy to use without requiring Python or command-line setup.
The current test release is v0.2.0-rc.3:
See GitHub Releases for all available versions.
The Linux DEB installer includes automatic desktop integration; see the installation guide for setup.
Installation and first launch: macOS · Windows · Linux. The guide also covers unsigned-app prompts on macOS and Windows.
- Drag a video into the application.
- Choose an output directory.
- Select an adaptive preset or fixed interval.
- Optionally mark one or more regions on the timeline.
- Choose JPEG or PNG.
- Click Start extraction.
For adaptive extraction, Low, Medium, and High control how densely keyframes are selected.
You can also navigate frame-by-frame and use Extract to save individual frames manually.
Manual captures are saved under manual_frames/<video-name>-<source-id>/.
Filenames use the original presentation timestamp (frame_pts_<timestamp>) to
distinguish variable-frame-rate frames; videos without timestamps use frame
numbers. Repeating a capture reuses its existing file. Older captures with
frame-number filenames remain in place and may be saved again under the new name.
Each run creates a timestamped directory containing the extracted images and metadata:
20260827_120000/
config.yaml
keyframes.csv
summary.txt
keyframes/
keyframe_0000_000000.jpg
keyframe_0001_000037.jpg
...
A CLI is included for scripted workflows:
./build/release/frame-extractor input.mp4 --output-dir outputRun the following for all available options:
./build/release/frame-extractor --helpFrame Extractor uses C++20, CMake, FFmpeg, OpenCV, SDL3, Dear ImGui, yaml-cpp, and Catch2.
See the development guide for complete build instructions for macOS, Windows, and Ubuntu Linux. Maintainer packaging and release details are in the release guide.
- Blur/sharpness-aware automatic selection is not yet implemented.
- macOS packages currently target Apple Silicon.
- Windows and Linux have received less hands-on testing than macOS.
Frame Extractor is released under the MIT License. Distributed packages also contain third-party software covered by the third-party notices.
Bug reports and contributions are welcome. See CONTRIBUTING.md. Please report suspected security vulnerabilities privately as described in SECURITY.md.