2021-01-08

New Features

If you select a specific element from your dataset or inference set, this new feature allows you to find similar, existing elements from that same set.

Collection Campaign

Similar to the above, this new workflow (pre-alpha) allows you to identify and collect new, unlabeled data samples that are most similar to the elements in an existing issue.

Reach out to us if you are interested in trying this out!

Multi-label Classification Flow

For classification tasks, we now support workflows where there may be multiple classification labels associated with a given frame. (Note: this is distinct from detection because each label applies to the whole frame).

The app accounts for this in the main grid view:

As well as in the confusion matrix:

More context on the above: our default confusion matrix works by associating each ground truth label with at most one inference --- with multiple labels, this paradigm is less well-defined, so we only calculate:

  • diagonal values --- matches between ground truth and inference

  • bottom row and right column --- (anything <> background) can be defined as any time a class is in the labels but not the inference, or vice-versa

To use this workflow, simply upload a new project with the primary task MULTI_LABEL_CLASSIFICATION.

App Improvements

Confusion Matrix Visualization for Many Classes

If your project supports more than 25 classes, you might have found the confusion matrix hard to decipher.

Now we have a chart-based view that should make it easier to search for the classes/confusions of interest:

Improved Classification Flow

When dealing with a classification task, you have a 1:1 relationship between frames and crops and the distinction is not necessary to keep in mind.

We've simplified the UI for classification workflows (project with primary task CLASSIFICATION ) to take this into account, so that there is only one tab (Crops) in the Embedding View:

Now you can search with new metrics-based filters (e.g. f1, precision, recall, support, false positives, false negatives, true positives, total confusions).

Metrics filters are associated with a given IOU and confidence threshold (default 50 and 0, respectively). If no classification is specified, weighted avg is used.

Now you can search for issues by name, state, and/or reporter:

Now you can search for projects by name! Still in the works: allowing you to search by dataset name and other project metadata.

Error Reporting for Image Uploads

Now, if your images aren't showing up in the app, a more informative error will show:

Last updated