Another great week of development! This week we’ve been concentrating on improving Images and sampling methods and a search! Details:
We’ve added two new labeling tags for images, key points, and polygons. That required redoing an image component from the ground up, but now it’s more robust and responsive.
Here are a few examples of how you can use new tags:
- keypoints for face annotation
<View> <Image name="img" value="$image" zoom="true"></Image> <KeyPointLabels name="tag" toName="img" strokewidth="5" fillcolor="red"> <Label value="Nose" background="yellow"></Label> <Label value="Left Eye" background="blue"></Label> <Label value="Right Eye" background="green"></Label> </KeyPointLabels> </View>
<View> <Image name="img" value="$image" showMousePos="true" zoom="true"></Image> <PolygonLabels name="tag" toName="img" strokewidth="5" fillcolor="red" pointstyle="circle" pointsize="small"> <Label value="Sign" background="red"></Label> <Label value="Road" background="blue"></Label> </PolygonLabels> </View>
If you specify
zoom="true" in your config, you get mouse wheel (or touchpad) zoom available on your images. That enables you to do pixel-perfect annotation!
Additional configuration for the sampling. When you go into the project settings, you can configure the way sampling of the tasks works. There are three different way right now:
- Sequential picks items in a sequence can be helpful when you upload the data in a specific order, and you want to label it in the same order.
- Random randomly picking the items from the dataset
- Active Learning use model and our extensive algorithms to pick the items for you to label (the hardest examples first), and as a result, you can label entire dataset faster.
Along with that, you get to control to show Ground Truth items first. That way you can validate the annotators (or your instruction) from the beginning.
The initial version of the search functionality made it into the UI. Check it out in the data manager!
New analytics doughnuts for RectangleLabels and Ground truths (#873)
Ordering for the agreements and prediction scores with null is fixed. (#888)
Now you have a better experience ordering items in the data manager.
Fixes for the ground truth items.
Datasets: support for ground truth added. (#871)
Get rid of the retraining period. If you connect one of the Heartex models (which you can do through project settings), it starts training right away.
Fail to retrain on low-agreement tasks
Part of Collaborator settings got moved to the Labeling tab. (#868)