Natural Language Search, Scale Model Zoo, Send For Re-Annotation, Send LidarAnnotation Tasks, Faster Semantic Segmentation, Display Settings Update, Faster Item Selection

Natural Language Search

• You can now search your image datasets in Nucleus with natural language queries, powered by CLIP embeddings. Simply describe what you’re looking for in English in the search bar, and quickly find relevant data. * Natural language search supports complex and nuanced queries, like “a pedestrian wearing a hat” or “a chaotic intersection” or “a blurry photo” for a self-driving car dataset. Try it out!
• Refine your natural language search results further by creating an Autotag. The combination of natural language search and Autotag refinement is a powerful workflow for targeted mining of rare data.

Scale Model Zoo

• We’ve released the Scale Model Zoo, a collection of pre-trained models that can be run on your Nucleus datasets.
• These models can be used to detect or segment common classes, which can add powerful metadata and evaluation metrics to find issues with your model or ground truth. Model Zoo inference can be run on-demand, or configured to happen automatically on data upload.
• The Nucleus Model Zoo currently supports MSCOCO-trained object detection and semantic segmentation as well as ImageNet-trained image classification.
• For a walkthrough of how to trigger a model zoo inference run, check out the Model Zoo guide at: https://dashboard.scale.com/nucleus/learning-guides

Submit Labeled Data for Re-Annotation

• Nucleus customers can now submit data items that already have ground truth associated with them for re-annotation through Scale.
• Existing image annotations can automatically be loaded as pre-labels ( “hypotheses”) for re-annotation, which can accelerate annotation time.

Submit LidarAnnotation Tasks to Projects from Nucleus

• Nucleus now supports the direct submission of lidar tasks through the GUI
• Users with 3D datasets that use Nucleus as tool to curate data for annotation are now able to also directly submit their pointclouds to labeling

Faster Semantic Segmentation

• We have improved Nucleus performance for all datasets with semantic segmentation tasks to ensure a seamless experience for users with segmentation use cases

Display Settings Update

• Nucleus now enables users to hide model predictions below a confidence threshold
• Users can now customize which types of annotations are shown or hidden when searching their datasets

Faster Item Selection

• We rewrote the Nucleus infinite scroll grid component for higher performance and lower memory consumption. Item selection, scrolling, and overall application FPS in the grid view is now faster, especially for datasets with large numbers of annotation per item.