Encord
Encord Active is a tool for machine learning and computer vision developers. It primarily focuses on model evaluation, data curation and active learning. This t...
Last verified:
What is Encord?
Encord Active is an active learning toolkit designed to help machine learning teams evaluate models, validate data and labels, and improve overall performance in computer vision projects. It enables users to visualize data, surface model failure modes, detect label errors automatically using vector embeddings and AI-assisted metrics, and prioritize high-value data for relabeling. Tightly integrated with Encord Annotate, it supports seamless workflows from data curation to annotation and model deployment, making it ideal for ML engineers and data scientists building production AI.
Key features include advanced model evaluation for robustness checks, data and label validation to build balanced datasets, similarity and natural language search, off-the-shelf and custom quality metrics, outlier and duplication detection, and collections for grouping data units. It handles images (jpg, png, tiff) and videos (mp4) up to 4K resolution, with support for labels like classification, bounding boxes, polygons, polylines, bitmasks, and keypoints. Users can compare model predictions against ground truth, export explainability reports, and integrate human-in-the-loop processes for iterative improvement.
Encord Active is for teams at any stage of their ML journey, from initial data collection to production models, particularly those working on multimodal data for physical AI applications like autonomous vehicles or medical imaging. It accelerates development by pre-empting blind spots, fixing underperforming models, and ensuring data integrity, as evidenced by customer gains like 67% increase in edge-case performance and 20% mAP boost.
The tool fosters collaboration with tagging, filtering, and sharing capabilities, while its open-source roots provide flexibility, though full features require an Encord Annotate account. It stands out for end-to-end active learning, reducing deployment timelines and enhancing model reliability in real-world scenarios.
Encord pricing
Pricing model: Free
Encord Active is available as a free tier with core features including up to 500,000 data units per project, unlimited projects, data exploration, label error detection, model evaluation, and integration with Annotate; full platform access including Annotate requires paid Encord plans (pricing not detailed on Active page, contact sales for enterprise options).
Encord pros
- Automatically detects label errors without manual review
- Supports natural language search for visual data
- Vector embeddings for similarity search
- Off-the-shelf and custom quality metrics
- Model evaluation with prediction import
- Outlier and image duplication detection
- Collections for grouping data units
- Integration with Encord Annotate
- Handles up to 500,000 data units per project
- Supports videos up to 2 hours at 30fps
- Robustness checks for data drift and blind spots
- Explainability reports for failure modes
- Balanced dataset building via filtering
- Nested attributes and custom metadata support
- Human-in-the-loop active learning workflows
- 67% edge-case performance gains reported
- 20% mAP increase from data curation
Encord cons
- Requires Encord Annotate account for use
- Limited to 500,000 data units per project
- Videos capped at 2 hours @ 30fps
- Performance drops over 4K resolution
- Recommends downscaling high-res media
- No support for data types beyond jpg/png/tiff/mp4
- Objects and classifications can't mix in model eval
- Limited video length for long recordings
- Hosted only by Encord platform
Frequently asked questions about Encord
How do I use Encord Active for active learning?
Encord Active supports active learning by letting you explore data to select labeling priorities, use acquisition functions for automatic selection, find label errors harming model performance, send data to Annotate for labeling, decompose model performance for focus areas, and tag subsets for test sets on edge cases.
What data types does Encord Active support?
It supports images in jpg, png, tiff formats and mp4 videos up to 4K resolution optimally; performance affects higher resolutions, so downscaling is recommended; labels include classification, bounding box, polygon, polyline, bitmask, key-point.
Can I import my model predictions into Encord Active?
Yes, import predictions via the Active UI or Active API; supports analysis on predictions compared to ground truth for evaluation.
Can I use Encord Active without an Encord account?
No, you need an Encord Annotate account as Active is tightly integrated and hosted by Encord.
What are Collections in Encord Active?
Collections save interesting groups of data units and labels to support downstream workflows like data curation, label correction, and model optimization.
Does it support natural language search?
Yes, natural language search allows querying visual data including images, videos, DICOM, labels, and metadata using everyday language.
How does it detect label errors?
Uses vector embeddings, AI-assisted quality metrics, model predictions to automatically surface problematic samples and common issue types.
What is the limit on video length?
Supports up to 2 hours of video at 30fps per project; longer videos may require splitting.
Can I compare multiple models?
Yes, integrate humans-in-the-loop to compare model performance iteratively and refine via active learning workflows.
Is data stored in my cloud storage?
Yes, if data is in AWS, GCP, Azure, or OTC cloud storage, it remains there; Encord Active analyzes without moving files.