Encord Active is available in two versions: Encord Active Cloud and Encord Active OS. Active Cloud is tightly integrated with Encord Annotate, with Active Cloud and Annotate being hosted by Encord. Encord Active OS is an open source toolkit that can be installed on a local computer/server.
Encord Active Cloud and Encord Active OS (open source) are active learning solutions that help you find failure modes in your models, and improve your data quality and model performance.
Use Active Cloud and Active OS to visualize your data, evaluate your models, surface model failure modes, find labeling mistakes, prioritize high-value data for relabeling and more!
Encord Active helps you understand and improve your data, labels, and models at all stages of your computer vision journey.
Whether you've just started collecting data, labeled your first batch of samples, or have multiple models in production, Encord Active can help you.
To give you a better idea about how Active Cloud and Annotate work together, here are some use cases.
Before going any further, you should know what a Collection is in Encord Active Cloud. Collections provide a way to save interesting groups of data units and labels, to support and guide your downstream workflow. For more information on Collections go here.
Alex, a DataOps manager at self-dr-AI-ving, faces challenges in managing and curating data for self-driving cars. Alex's team struggles with scattered data, overwhelming amounts of data, unclear workflows, and an inefficient data curation processes. Alex is currently a big user of Encord Annotate, but would like to provide better datasets for annotation.
Initial setup: Alex gathers a large number of images and gets them imported into Active. Alex then logs into Encord and navigates to Active (freemium).
First collection: Alex opens the Project and after searching, sorting, and filtering the data, selects the images and clicks Add to a Collection and then clicks New Collection. Alex names the Collection RoadSigns as the Collection is designed for annotating road signs for the team.
Data curation: Alex then further bulk-finds traffic sign images using the embeddings and similarity search. Alex then clicks Add to a Collection and then clicks Existing Collection and adds them images to the RoadSigns Collection in a matter of clicks.
Labeling workflow: Thinking about different use-cases (for example, "Labeling" and "Data quality") Alex assigns good quality road signs for labeling, and bad quality road signs for "Data quality" and future "Deletion". In the future Alex might use "Active learning" to prioritize the data for labeling.
Sent to Annotate: Alex goes to the Collections page, selects the Roadsigns Collection and clicks Create Dataset. Active creates the dataset and a new project in Annotate. Alex then configures the workflow, annotators, and reviewers for the Project in Annotate.
Review and insights: At the end of the week, Alex reviews the RoadSigns Project in Annotate. The dataset has been annotated. Alex goes to Active, clicks the More button on the Project then clicks Sync Project Data. Alex then clicks the Project, clicks Analytics and then Model Predictions where Alex gains insights into:
- Number of labels per image
- Quality of annotations
- Distribution of annotations across metric
The process is seamless and fast, and Alex can focus on more strategic tasks while her team enjoys a much-improved, streamlined data curation workflow.
Chris, a Machine Learning Engineer at a micro-mobility startup, has been working with Encord Annotate. His team is dealing with a large set of scooter images that need accurate labeling for obstacle detection. After an initial round of annotations, Chris notices that some labels are incorrect or ambiguous. This has a significant impact on the performance of the ML model.
Access Collections: Chris logs into Encord Active. Chris opens the Scooters project that was imported from Annotate. Chris goes to the Collections page for the project. Chris browses the existing Collections to see if he should create a new Collection or add to an existing Collection.
Data exploration: Chris searches, sorts, and filters the previously annotated scooter images and identifies those that need re-labelling.
Create re-labeling Collection: Chris selects the images and clicks Add to a Collection. Chris then clicks New Collection and names the Collection Re-label - Scooters.
Initiate re-labelling: With the Collection ready, Chris returns to the Collections page, selects the Re-label - Scooters Collections and clicks Create Dataset. The Collection is sent to Annotate.
Assigning in Annotate: In Annotate, Chris assigns re-labelling tasks to specific annotators. Annotators then complete their relabelling tasks.
Quality check in Active: After the re-labeling tasks are completed, Chris clicks Sync Project Data. The updated labels then sync back to the Project. Chris reviews the changes, confirms the label quality, and plans for model re-training.
The "Collections" feature has simplified the task of identifying and re-labeling inaccurate or ambiguous data, streamlining the entire data annotation and quality control process for Chris and his team.
|Features||Active Open Source||Active Cloud|
|Number of images||25,000 per project (unlimited projects)||200,000 per project (unlimited projects)|
|Videos||-||2 hours @ 30fps|
|Off-the-shelf quality metrics||✅||✅|
|Custom quality metrics||✅||✅|
|Data and label tagging||✅||✅|
|Image duplication detection||✅||✅|
|Label error detection||✅||✅|
|Natural language search||-||✅|
|Search by image||-||✅|
|Integration with Encord Annotate||-||✅|
1: Objects and classifications are both supported, but cannot be mixed. Classification support currently only includes support for a single radio button.
Updated 15 days ago