Open source application to help manually label video datasets for use in machine learning
range-tagger is a Qt-based desktop application designed to help you quickly label sections of long videos manually. It is designed to help you assign every frame in your video with a label so you can then use the data to train machine learning algorithms. range-tagger is still in alpha but is appropriate for light usage today.
You can learn more about the motivation and design of range-tagger here.
When developing machine learning models, having properly labeled data for tranining is very important. Let's suppose you want to build a machine learning model to automatically classify parts of a television broadcast into two classes: "commercial" or "part of the real show." Your first step is going to be collecting a large amount of video data, but then you will need to manually label a portion of it for model training purposes. range-tagger would let you quickly and accurately label segments of your videos as "commercials" or "show." You would produce a CSV that looks something like this:
| Start Frame | End Frame | Label |
|---|---|---|
| 0 | 1205 | commercial |
| 1206 | 7660 | show |
| 7661 | 9122 | commercial |
| 9123 | 14565 | show |
| ... | ... | ... |
which you could then link with image frames on disk that you feed into your model.
range-tagger does not handle "object detection" use cases; you can use a tool like VoTT to handle object detection. range-tagger took inspiration from VoTT for its design, but ultimately handles a different use case.
You can find the latest binary release here, and you can find a demo video here. Right now, range-tagger is only available on OSX; this is due to a single dependency. If you would like to use range-tagger on Windows or Linux, just post an issue and I can help create a build for your platform.
range-tagger uses keyboard shortcuts to help make your life manually labeling videos much simpler. The shortcuts are fairly quick to learn and will help you multiply your labeling productivity:
| Shortcut | Purpose |
|---|---|
| ⇧⌘C | Create a new segment at the current frame |
| ⇧⌘E | End the selected segment at the current frame |
| ⌘Z | Undo |
| ⇧⌘Z | Redo |
| ⌘1 | Label current segment with the first tag |
| ⌘2 | Label current segment with the second tag |
| ⌘3 | Label current segment with the third tag |
| ⌘4 | Label current segment with the fourth tag |
| a | Move backward 100 times the base number of frames |
| s | Move backward 10 times the base number of frames |
| d | Move backward 1 times the base number of frames |
| f | Move backward a single frame |
| j | Move forward a single frame |
| k | Move forward 1 times the base number of frames |
| l | Move forward 10 times the base number of frames |
| ; | Move forward 10 times the base number of frames |
And these shortcuts are shown extensively in the demo video here.