Home / Products / RATATA
RATATA
The name is the method. Building a computer vision model is never one pass. It is a loop, walked many times: record, annotate, train, and again. RATATA is the platform to make that loop fast and traceable, so the work that matters stays on your data and your models.
What problem RATATA solves
Most vision projects don’t fail on model architecture. They fail on everything around it. Fragmented annotations, untracked model iterations, and no clear way to find the high-impact images actually worth labeling lead to wasted effort.
RATATA is an end-to-end platform for the complete dataset lifecycle. It unifies capture, storage, inspection, annotation, inference, comparison, and export so you can focus your labeling effort exactly where it drives model performance.
Finding the right images
A critical step in model improvement is narrowing a massive dataset down to the few images that matter right now.
RATATA automatically builds dynamic filters from your dataset. It inspects what is actually present and offers the relevant controls.
- Image metadata: camera, lens, exposure, gain, capture time, altitude
- Annotation: by type and by label
- Predictions: by model, type, label, and a confidence range
- Context attributes: free-form model-produced attributes, filtered by value
- Batch metadata: whatever was recorded about the capture session
Filters combine effortlessly. A complex query—such as “Show every image where the model predicts litter with 0.4–0.6 confidence, captured after sunset, in Batch X” takes just a few clicks. The result instantly syncs across the entire interface, narrowing your gallery, map, statistics, and annotation queue simultaneously.
Geography as a first-class dimension
Map inspection
For field data, where is often the question. Geolocated images appear on a map, and the map does more than display them.
- Draw a polygon on the map to filter the dataset spatially. Everything else updates to match.
- Density heatmap answers where there is a lot of something.
- Interpolation surface estimates a continuous surface between sparse sample points and clips it to a field boundary.
Both can be driven from ground-truth annotations or from model predictions, and weighted by detection count, coverage area, or coverage fraction of the image.
The loop · Record
Record
Images arrive from wherever they are captured: field cameras, drones, vehicle-mounted rigs, handheld, or a folder of existing material uploaded through the browser. On image upload, RATATA reads the metadata, camera make and model, lens, exposure time, f-number, gain, GPS position, and capture time, etc., and stores it on the image record. That metadata is filterable immediately, which is what makes the later steps fast.
RATATA supports both RAW images, and developed images.
The loop · Annotate
Annotate
Annotation happens in the browser. RATATA supports different annotation types for various use cases.
You can use existing AI models to pre-annotate images, or to filter to only show images, where a model has detected a given object with a certain confidence range, to focus your annotation time. This is especially valuable in datasets with rare occurrences of a given class.
| Type | Use |
|---|---|
| rectangle | Standard object detection |
| polygon | Irregular shapes, precise extents |
| points | Keypoints and skeletons |
| mask / semantic_segmentation | Per-pixel coverage |
| tag | Whole-image classification |
| negative | Confirmed absence, as important to training as presence |
Keypoint work is driven by reusable skeleton templates. Define the named keypoints and the links between them once per project, in a visual editor, on a normalised canvas. The template then stamps onto any image regardless of resolution. Annotations can also carry a track ID, so the same object followed across frames stays one identity.
The loop · Train
Train
Registered models run against your data from inside the platform. Run one model on one image and get the prediction back immediately, or enqueue the entire filtered set, hundreds of thousands of images if that is what you selected.
The loop · Annotate, train, again
RATATATATA
After the first model run, you need to know where the model is wrong, and you need to fix exactly those cases.
RATATA’s Objects view puts predictions and ground truth against each other directly. Choose a model, set an IoU threshold, and filter to disagreements only: the cases where the model and the annotator do not match. Sort by confidence to surface confident mistakes first, since those are the ones worth your attention. Correct them, or bulk-remap labels across a whole selection in one operation, and the corrected set feeds the next training round.
That is the second TA in the name. And the third, and the fourth…
From the field, live
Fleet management
RATATA also manages the cameras. Registered AI Lab devices appear in a fleet view with their type, status and time since last contact, backed by ThingsBoard for device telemetry and management.
When you need to see what a device sees right now, you can open a live video stream. Useful for verifying a deployment without driving to it, and for checking that a camera is pointed where you think it is.
Built for teams and real data governance
Group and permission control
Access is controlled by group membership, to allow collaboration and distribute annotation effort, and you can grant viewer access, which allows a user to view, but not edit data.
Everything is available through a documented REST API, so RATATA can sit behind your own tooling as the system of record.