Five Services.
One Quality Standard.
Aeterna Data specializes in computer vision annotation and visual AI evaluation. We use a consistent workflow and follow the IAA quality standard, measuring and reporting results for each batch.
Bounding Box Annotation
We deliver accurate object detection labeling for production-ready computer vision models.
Bounding box annotation is the foundation of object detection training. At Aeterna Data, we label each object with clear, class-specific guidelines. This ensures that every annotator follows the same rules for each object type across the dataset.
We offer standard output formats like COCO JSON and YOLO TXT. Every batch comes with an IAA quality report that shows our annotators' Cohen's Kappa scores.
Image Segmentation
We identify objects down to the pixel, giving you accurate training data for your visual AI projects.
Image segmentation needs more precision than bounding box annotation because every pixel must be assigned to the correct class or instance. Aeterna Data uses polygon annotation and checks quality for each class to keep boundaries accurate, even in complex scenes.
We work with both semantic segmentation, where each pixel in the image is classified, and instance segmentation, where we mask each object separately. The pilot phase helps us set a baseline for annotation time and quality before starting production.
Image Classification
We organized the visual dataset using a clear structure and a consistent taxonomy.
Image classification means giving each image in a dataset one or more category labels. The main challenge is keeping the taxonomy consistent. Aeterna Data addresses this by using clear labeling guidelines for each class and checking annotator decisions to ensure categories stay consistent across large datasets.
We handle both single-label and multi-label classification tasks. Before starting production, we run a pilot phase to check the taxonomy and find any unclear category boundaries that should be clarified in the annotation guidelines.
Visual RLHF Evaluation
We train visual AI reward models using data that reflects human preferences.
In visual RLHF evaluation, annotators compare AI-generated images and give structured preference judgments. Instead of simply choosing a favorite, they assess images based on specific qualities like accuracy, coherence, and safety. Aeterna Data uses clear rubrics to make sure annotator judgments stay consistent and measurable.
We offer pairwise preference ranking, absolute quality rating, and safety classification for image generation models. We also track rubric drift across batches to make sure the reward model training gets a consistent signal throughout the dataset.
Dataset Validation & QA
We catch annotation errors before your model can learn from them.
Dataset validation checks annotated datasets for label consistency, class distribution problems, and systematic annotation mistakes. Finding these issues before training saves money compared to discovering them only after a model performs poorly in evaluation.
Aeterna Data checks how much annotators agree on existing datasets using Cohen's Kappa and Fleiss' Kappa. It also finds error patterns for each annotator and class, then provides a clear quality report with specific suggestions for reannotation.
Inter-Annotator Agreement
All Aeterna Data services follow the same inter-annotator agreement standard. Whether it is bounding box, segmentation, classification, visual RLHF, or dataset QA, our quality threshold stays consistent across every service.
Pairwise agreement between any two annotators on the same task must meet or exceed the 0.80 threshold.
Multi-annotator agreement across three or more raters must meet the 0.75 threshold for full-team consistency.
Your Dataset. Your Tools. Our Team.
Aeterna Data works with the annotation tools your team already uses. You do not have to switch to a specific platform. We fit into your existing setup.
We begin every project with a pilot phase. This lets us set up the annotation workflow, review the guidelines, and check that the quality meets your standards before we start full production. By doing this, we can catch any workflow issues early, rather than after labeling thousands of images.
No platform lock-in
We work with CVAT, Label Studio, Roboflow, Scale, and can also use your own internal tool.
Pilot before production
We start every project with a fixed-scope pilot. This helps us check the workflow and make sure the quality meets your standards before we move to full production.
Formal agreements first
We sign an NDA and DPA before receiving any data from you. We recognize your training dataset is proprietary and handle it with the utmost care.
Five Ways to Work With Us
Fixed-scope trial to validate workflow and quality before committing.
Per-batch annotation work with IAA reports on every delivery.
Large-scale annotation engagements with dedicated capacity and volume pricing.
A named team assigned exclusively to your project with consistent quality output.
Ongoing annotation partner for teams with continuous dataset needs.
Not Sure Which Service You Need?
Please provide details about your dataset and model. Once we receive your information, we will recommend the most suitable service and send a pilot plan within 48 hours.