How to accelerate your computer vision pipeline with SuperAnnotate + 3LC

Building and refining high-quality datasets for computer vision can be time-consuming—especially once you begin training a model and discover labeling gaps, errors, or missing classes. In this post, we’ll walk through a streamlined, iterative workflow that combines the power of SuperAnnotate for annotation with 3LC for real-time model feedback and dataset insights. You’ll learn how to identify labeling issues, fix them quickly, and continuously improve your model’s performance.

Why iterative annotation & model feedback matter

Even the most robust annotation processes face challenges once a model starts training. Common issues include:

• Missing labels: Objects that should have been labeled were overlooked, leading to ambiguity for the model.
• Incorrect or inaccurate labels: Some objects were assigned to the wrong class or labeled imprecisely, adding noise and reducing the overall quality of the dataset.
• Need for additional classes: Certain object categories may be absent from the dataset. For instance, a model might misclassify unlabeled objects by assigning them to an incorrect existing class, signaling the need for additional categories to help the model separate similar features.
• Class imbalance: Even if a dataset is “balanced” by raw numbers, embedding space, and model feedback might tell a different story. 

3LC makes detecting these issues straightforward. It offers real-time training insights using per-sample metrics and embeddings directly from your model. Since these metrics come from your own model, they’re far more relevant than those generated externally. They expose labeling issues, allowing you to flag them for review in SuperAnnotate. This feedback loop significantly enhances dataset quality and ultimately improves model performance.

Step-by-step guide

Let’s walk through an example project: suppose we want to detect bees in images of flowerbeds. Our initial dataset contains around 7,000 images. 

Here’s how we can refine it iteratively using SuperAnnotate and 3LC:

Step 1: Annotate with SuperAnnotate

1. Create a project in SuperAnnotate:
   
   • Set up a new project in SuperAnnotate and upload your ~7,000 flowerbed images.
   • Define your classes (e.g., “Bee” vs. “Background Objects”).
2. Label your dataset:
   
   • Use SuperAnnotate’s efficient annotation tools to draw bounding boxes around each bee.
   • Assign the correct class to each bounding box.

Your initial annotations form the foundation for your computer vision model. However, as you’ll see in the next steps, you’ll revisit them once you have real-world feedback from the model.

Step 2: Import annotations into 3LC

After completing initial annotations, bring the labeled dataset into 3LC for model training and per-sample metric collection. Using 3LC’s Python library, tlc, you can import a dataset directly from SuperAnnotate. The only manual step is triggering an export in SuperAnnotate.

import tlc
from superannotate import SAClient
sa_client = SAClient(token="superannotate_token")


train_table = tlc.Table.from_superannotate(
	sa_client=sa_client,
	sa_project="Bees",
	sa_export="name_of_export.zip",
	images_folder="/location/of/images/on/disk",
	project_name="Bees",
	dataset_name="train"
)

This code snippet converts your annotated dataset into a 3LC-compatible format while preserving the bounding boxes and classes.

Step 3: Train a model & collect detailed metrics

With your dataset in 3LC, train a YOLO model (using the Ultralytics integration) and automatically gather per-sample metrics throughout the training process.

from ultralytics.utils.tlc import TLCYOLO

model = TLCYOLO("yolo11mpt")

tables = {
    "train": train_table,
    "val": val_table
)
model.train(tables=tables, epochs=50)

3LC collects and creates detailed information, including:

• Confidence and IOU scores for each detection.
• Embeddings for individual bounding boxes.
• False positives, false negatives, and other metrics at multiple training stages.

Step 4: Analyze in 3LC

Once training is completed, head to your 3LC dashboard. You’ll see visualizations that pinpoint potential dataset issues, such as:

• Missing or incorrect labels: Sort by the model’s highest-confidence false positives to find cases where a “Bee” wasn’t labeled.
• Need for additional classes: If the model keeps confusing certain black-and-yellow flowers for bees, introducing a new class for these flowers can help the model differentiate.
• Class imbalances or hard examples: Explore embeddings to spot underrepresented classes or especially challenging samples.

Example: You might discover ~100 images with high-confidence “Bee” detections that weren’t labeled in the training data. These missing bounding boxes become prime targets for annotation updates. With just a few clicks, you add all these new bounding boxes to your dataset as label suggestions.

Step 5: Export updated annotations back to SuperAnnotate

Once you’ve identified problem areas, you can generate label suggestions directly from 3LC with a few clicks:

1. Save 3LC table revision: Incorporate the label corrections or new bounding boxes.
2. Export to SuperAnnotate: Send these suggestions (added, edited, or removed bounding boxes) back to SuperAnnotate.

tlc.add_edits_to_superannotate_project(
table=train_table.latest(),
sa_client=sa_client,
sa_project="Bees",
)

Back in SuperAnnotate, your labeling team can review and confirm the suggested bounding boxes. This adds a second layer of quality control and ensures any automated corrections align with human judgment.

Step 6: Rinse and repeat

With your improved dataset, retrain the model in 3LC. Keep iterating until you’re satisfied with the model’s performance. This closed-loop process is faster and more accurate than manual error-spotting alone and continually drives higher-quality results.

Future outlook: Beyond computer vision

While 3LC started with computer vision, this iterative workflow can be extended to other AI modalities—such as language and audio—in the future. The synergy between annotation and detailed model insights remains the same, ensuring any dataset can be refined continuously as you train.

About SuperAnnotate

SuperAnnotate is the leading data annotation platform for creating and managing high-quality labeled datasets for AI applications. Trusted by top companies worldwide, it combines advanced tools with expert services to streamline the annotation process.

About 3LC

3LC leverages detailed model feedback during training to uncover actionable insights into datasets. Designed for seamless integration into your existing training workflow, 3LC enables teams to refine labels, adjust datasets, and visualize improvements without disruption.

Get started

Interested in trying out this joint workflow? Contact us or visit 3LC.ai to learn how you can integrate 3LC with SuperAnnotate. Accelerate your model development and continuously improve your datasets, resulting in more accurate models—faster than ever before!