Check labels
Finds images whose label is likely wrong, from the labels of the images that look most like them, and images unlike any other.
How it works
For each image, its Neighbours most similar images (by the embeddings of Embed images) vote with their labels (the folders), weighted by similarity. When the image's own label gets less than Agreement (%) of the vote and another label wins, the image is flagged label, with the label its neighbours suggest and how sure that is.
Outliers (%): the share of images least like their neighbours, flagged outlier (empty frames, odd shots); 0: none. A confusion table shows which labels are mixed up with which. Curate and split removes the flagged images, or a person reviews them first.
Inputs
| Input | Takes | What it is |
|---|---|---|
Embeddingsembeddings | Files npy, json | embeddings.npy and paths.json of Embed images |
Outputs
| Output | Gives | What it is |
|---|---|---|
Flagstable | Table csv | label_check.csv: one row per image (path, label, suggested, agreement, confidence, isolation, flag, reason); flag is empty when the image is fine |
Confusionconfusion | Table csv | label_confusion.csv: how many images of each label the neighbours put in each label |
Live telemetrytelemetry | Live telemetry | What the block sends while it runs, for views such as Chart or 3D view. |
Settings
| Setting | Default | What it does |
|---|---|---|
Neighboursneighbours |
10 | From 1 to 100 |
Agreement (%)agreement |
30 | Flagged when the image's own label gets less of the vote than this From 0 to 100 |
Outliers (%)outliers |
1 | From 0 to 20 |
Where it runs
On one of your computers with the EsaSim worker: Any worker.
The worker installs the Python packages it needs when they are missing: numpy.
In its code, settings are ctx.params["key"], inputs ctx.inputs["id"] and outputs ctx.output("id", "name.ext"); see Custom blocks.