mirror of
https://github.com/ultralytics/ultralytics
synced 2026-04-21 14:07:18 +00:00
docs: dataset health & clustering
This commit is contained in:
Sergiu Waxmann
parent
f4d0fda2cb
commit
775a8ff45f
2 changed files with 87 additions and 1 deletions
|
|
@ -218,7 +218,9 @@ graph LR
|
|||
|
||||
## Browse Images
|
||||
|
||||
View your dataset images in multiple layouts:
|
||||
View your dataset images in multiple layouts.
|
||||
|
||||
Switch to the [Health & Clustering](#health-clustering) panel to inspect image quality and explore clusters visually.
|
||||
|
||||
| View | Description |
|
||||
| ----------- | --------------------------------------------------------------------------------- |
|
||||
|
|
@ -285,6 +287,86 @@ Filter images by their dataset split:
|
|||
| **Val** | Used for validation during training |
|
||||
| **Test** | Used for final evaluation |
|
||||
|
||||
## Health & Clustering
|
||||
|
||||
The `Health & Clustering` panel analyzes your dataset for quality issues and visualizes image similarity as an interactive 2D scatter plot. Open it from the scatter-chart icon in the gallery toolbar on any dataset page.
|
||||
|
||||
<!-- TODO(screenshot): platform-datasets-health-clustering-empty-state.avif -->
|
||||
|
||||
|
||||
### Running Analysis
|
||||
|
||||
Dataset owners and editors can start an analysis:
|
||||
|
||||
1. Open a dataset and click the scatter-chart icon in the gallery toolbar
|
||||
2. Click `Analyze Dataset`
|
||||
3. Wait for the progress bar to finish — results appear in the same panel
|
||||
|
||||
Analysis runs in the background and can take a few minutes depending on the size of your dataset. You can close the panel or leave the page and come back later.
|
||||
|
||||
<!-- TODO(screenshot): platform-datasets-health-clustering-progress.avif -->
|
||||
|
||||
|
||||
!!! tip "Cancel & Retry"
|
||||
|
||||
If an analysis gets stuck, you will see a `Cancel & Retry` button to reset it and start over.
|
||||
|
||||
|
||||
### Quality Metrics
|
||||
|
||||
After analysis, the panel summarizes how many images fall into each issue category:
|
||||
|
||||
| Metric | What it finds |
|
||||
| ------------------ | -------------------------------------------------- |
|
||||
| **Blurry** | Images that are out of focus or motion-blurred |
|
||||
| **Overexposed** | Images that are too bright or washed out |
|
||||
| **Low contrast** | Images with flat tones and little visual variation |
|
||||
| **Near-duplicate** | Images that are visually identical or very similar |
|
||||
|
||||
Each metric also appears as a clickable filter badge on the scatter plot.
|
||||
|
||||
### Cluster Visualization
|
||||
|
||||
Once analysis completes, the panel shows a 2D scatter of all images. Similar-looking images appear close together, so duplicates, outliers, and under-represented visual subsets become visually obvious.
|
||||
|
||||
|
||||
|
||||
#### View
|
||||
|
||||
Change how data points are shaded with the `View` dropdown in the panel toolbar. Switch view modes at any time — the plot re-colors instantly so you can see how splits, classes, or image properties are distributed across your clusters:
|
||||
|
||||
| Option | Shading |
|
||||
| --------------- | ---------------------------------------------------- |
|
||||
| **Splits** | Train / Val / Test — spot split imbalance per region |
|
||||
| **Classes** | Annotation class on each image |
|
||||
| **Width** | Image width |
|
||||
| **Height** | Image height |
|
||||
| **Size** | File size |
|
||||
| **Annotations** | Number of annotations per image |
|
||||
|
||||
Categorical modes (Splits, Classes) render a legend; numeric modes render a gradient scale.
|
||||
|
||||
<!-- TODO(screenshot): platform-datasets-health-clustering-view-modes.avif -->
|
||||
|
||||
|
||||
#### Quality Filters
|
||||
|
||||
Click any of the available quality badges to keep only matching points on the plot and in the gallery below. Filters compose — enabling Blurry and Near-duplicate shows both sets at once.
|
||||
|
||||
#### Selecting Images
|
||||
|
||||
Selected points immediately filter the gallery below the scatter plot, so you can brush a suspicious cluster and inspect, relabel, move, or delete the matching images using the usual [image operations](#image-operations).
|
||||
|
||||
<!-- TODO(screenshot): platform-datasets-health-clustering-brush-selection.avif -->
|
||||
|
||||
|
||||
### Re-analysis
|
||||
|
||||
If you add or remove images, re-run the analysis from the panel header:
|
||||
|
||||
- **Update Analysis** — appears when your dataset has changed significantly since the last run
|
||||
- **Re-analyze** — recompute at any time
|
||||
|
||||
## Dataset Tabs
|
||||
|
||||
Each dataset page can show up to six tabs, depending on the dataset state and your permissions:
|
||||
|
|
|
|||
|
|
@ -127,6 +127,10 @@ The `Charts` tab provides automatic analysis including:
|
|||
- **Annotation Locations**: 2D heatmap of bounding box center positions
|
||||
- **Image Dimensions**: 2D heatmap of width vs height with aspect ratio guide lines
|
||||
|
||||
### Health and Clustering
|
||||
|
||||
Explore your dataset as an interactive 2D scatter plot, detect blurry, overexposed, low-contrast, and near-duplicate images. Brush a region of the plot to filter the gallery by cluster. See [Health & Clustering](datasets.md#health-clustering) for details.
|
||||
|
||||
## Quick Links
|
||||
|
||||
- [**Datasets**](datasets.md): Upload, manage, and export your training data
|
||||
|
|
|
|||
Loading…
Reference in a new issue