3LC works across CV tasks
Classification
- Find samples which are misclassified by your model
- See per-sample metrics evolve throughout training
- Edit your labels and instantly re-train
Object Detection
- Add, remove, resize or relabel any bounding box
- Filter on per bounding box properties like confidence or IoU
- Apply thousands of edits with a single click based on model
predictions
Pose Estimation
- Visualize and edit keypoints, skeleton lines, and bounding boxes as a unified pose annotation
- Derive per-keypoint and per-pose OKS (Object Keypoint Similarity) metrics to pinpoint where your model struggles
- Import from COCO or YOLO keypoint formats, with full training support via Ultralytics and SuperGradients
Oriented Bounding Boxes (OBB)
- Detect rotated objects with tightly-fitting oriented rectangles defined by center, size, and angle
- Import from YOLO-OBB format or create custom OBB tables with full Ultralytics training integration
- Add, remove, resize, rotate, or relabel any oriented bounding box in the Dashboard
Semantic Segmentation
- View and edit class masks at the per pixel level
- Identify and alleviate class and object-size imbalances across your segmentation labels
- Use per- class IoU metrics to understand exactly where your model is struggling
Instance Segmentation
- View and edit per-instance segmentation masks with pixel-level precision
- Convert predicted masks to ground truth with a single click, or use SAM to auto-generate masks
- Import from COCO or YOLO segmentation formats and scale datasets rapidly with active labeling workflows
3LC integrates easily into your code
Feature Comparison Matrix
| Feature / Capability | 3LC | Other Data Curation Tools |
|---|---|---|
| Per-sample metrics collection Per-sample metrics collection across epochs to track model performance | Yes | No |
| Human-in-the-loop, semi-automated labeling Light-touch, human-in-the-loop workflow combining our data-centric AI platform, expert guidance, and iterative Vision Language Model fine-tuning to outperform SME-only labeling—delivering higher accuracy in a fraction of the time and cost | Yes | No |
| Instant Feedback loop between data revisions and retraining Instant feedback loops for rapid model retraining after data updates | Yes | No |
| Dataset Versioning Non-destructive edits to individual samples, maintaining a git-like lineage of changes in your dataset revisions | Yes | Limited (Snapshots) |
| Rich Data Visualization Tools Up to 6D visualization | Yes | Yes |
| Insightful Performance Clustering Performance clustering to diagnose and identify where your model is struggling so you can take meaningful action | Yes | Limited |
| Robust Data Handling Handles data >100k images | Yes | Yes |
| Filter any data, at any level Filter datasets and metrics by any property. Combine conditions from sample-level attributes to detailed sub-sample properties. Quickly focus on the most relevant subsets to target exactly the data you need. | Yes | Limited (views & plugins) |
| Synthetic Data Analysis Ability to evaluate how specific synthetic data samples affect model performance | Yes | Limited |
| Complete Label Editing Capabilities Targeted label and data editing capabilities to achieve optimal model performance | Yes | No |
| Generate Smaller Iimpactful Datasets Automatically generate smaller, high-impact datasets by detecting and removing low-value or underperforming samples | Yes | No |
| Bulk Label Reassignment Automated Bulk Label Reassignment — saving teams significant time and effort | Yes | No |
| Custom Model Integration Custom integration with your existing Python Code, integrates with PyTorch, Tensorflow and other ML Frameworks. Pre-built integrations with popular tools and ecosystems, such as YOLO, Detectron2, Hugging Face and PyTorch Lightning. | Yes | Limited (through plugin APIs) |
| Scalable Deployments Scalable Deployment for On-Prem, Cloud, or Hybrid Environments | Yes | Yes |
Human in the Loop
Improve your data + model
Data
Debugger
- Find important or inefficient samples
- Understand what samples work and where your model struggles
- Improve your model by changing your data
Dataset
Versioning
- Make sparse, non-destructive edits to individual samples or in a batch
- Maintain a lineage of all changes and restore any previous revisions
- Avoid duplicating data
Experiment
Tracking
- Aggregate metrics by sample features, rather than just epoch, to spot hidden trends
- Tie each training run to a specific dataset revision for full reproducibility
- Find the best combination of model hyperparameters and dataset modifications