Question 1

What is Payana?

Accepted Answer

Payana is the managed vision intelligence service by Baaz. We turn raw images, videos, and wearable-camera footage into training-ready datasets: AI-assisted labeling, human review, dataset versioning with train/valid/test splits, and export to COCO JSON, YOLO, Pascal VOC, PNG masks, LeRobot, RLDS, Parquet and JSONL.

Question 2

How is Payana different from self-serve annotation tools?

Accepted Answer

Annotation tools give you an editor - your team still spends weeks drawing boxes and polygons. Payana is a service: Baaz's annotation team and AI pipeline do the labeling for you, quality-review every dataset, and deliver versioned exports. You go from raw footage to a training-ready dataset without hiring or training annotators.

Question 3

What annotation types does Payana support?

Accepted Answer

Bounding box annotation for object detection, polygon annotation for instance segmentation, pixel-level masks for semantic segmentation, keypoint and skeleton annotation for pose estimation, and single- or multi-label image classification. Each project type maps automatically to the correct export formats.

Question 4

How does AI-assisted labeling work in Payana?

Accepted Answer

Click an object or drag a rough box and our GPU segmentation engine finds the exact object boundary - a clean, editable polygon in about a second. Existing rough polygons can be refined the same way, recovering parts a manual outline missed. AI proposes, humans verify: that combination is what keeps both speed and accuracy high.

Question 5

Which dataset export formats are available?

Accepted Answer

Computer vision: COCO JSON, YOLO TXT with data.yaml (detection, segmentation and pose variants), Pascal VOC XML, PNG class-index masks, and classification folder/CSV layouts. Robot learning: LeRobot v3, RLDS/Open X-Embodiment style episodes, Parquet, and JSONL. Every export ships as a zip with train/valid/test splits, ready for PyTorch, TensorFlow or Ultralytics training pipelines.

Question 6

Can Payana turn video into a labeled dataset?

Accepted Answer

Yes, two ways. Any video can be sampled into frames (for example one frame per second) for image labeling. Wearable or head-mounted footage can also flow through Payana's automatic episode pipeline, where a 7-model GPU stack - a vision-language model, object and open-vocabulary detection, pixel-accurate segmentation, speech transcription, on-task scoring and hand-pose estimation - annotates every second without manual work.

Question 7

What are robot-learning datasets, and does Payana support LeRobot?

Accepted Answer

Robot-learning (imitation learning) datasets pair what a person sees with what their hands do, so robots can learn manipulation from human demonstrations. Payana captures egocentric video, extracts per-frame actions including a 7-dimensional end-effector state, and exports LeRobot v3 and RLDS/Open X-Embodiment style episodes - purpose-built for robotics foundation models.

Question 8

How is annotation quality assured?

Accepted Answer

Three layers: AI pre-annotation with confidence scores, human review of every episode and image in a built-in review workspace, and automatic curation that flags off-task or low-quality segments before they enter a dataset. Datasets ship with quality metrics, and reviewer overrides are tracked.

Question 9

Which industries does Payana serve?

Accepted Answer

Manufacturing and assembly-line analytics, warehouse and logistics robotics, healthcare and medical imaging annotation, agriculture and field robotics, retail and inventory vision, and any team building computer-vision models or training robots from human demonstrations.

Question 10

Can I use Payana datasets to train YOLO or other models directly?

Accepted Answer

Yes. The YOLO export unzips into the exact folder structure Ultralytics expects - images, labels and data.yaml - so training starts with a single command. COCO exports drop into standard detectron/mmdetection-style loaders, and classification exports load with PyTorch ImageFolder unchanged.

Question 11

Can Payana train a model for me, not just deliver the dataset?

Accepted Answer

Yes. Model training is part of the service: pick a labeled dataset - a Payana-labeled project, an annotated episode dataset, or a dataset zip you upload - choose a model size and training length, and we train a detection, segmentation, keypoint or classification model on our GPUs. You watch live progress, get honest accuracy metrics measured on held-out validation data, and download deployable model weights when it finishes. You can also retrain with more data later, continue from a previous training run, or fine-tune a model you already have.

Question 12

How fast is the turnaround?

Accepted Answer

AI pre-annotation runs in minutes on GPU - a one-minute video is fully machine-annotated shortly after upload. End-to-end delivery including human review is typically days, not weeks, depending on volume; the intro call gives you a concrete estimate for your dataset size.

Question 13

Where is my data stored and how is it secured?

Accepted Answer

On dedicated private infrastructure managed by Baaz: access-controlled storage, isolated processing environments, and short-lived, time-limited download links. Your images and videos are never used to train third-party models and never appear in a public dataset pool.

Question 14

How do I get started with Payana?

Accepted Answer

Book a 30-minute intro call with the Baaz team. We scope your use case - objects, environments, formats - then handle capture, annotation, review and delivery end to end. You receive versioned, training-ready datasets without hiring an annotation team.

Capability	Payana	Self-serve tools	Labeling vendors	Data collection vendors
Data collection done for you (wearable / field capture)Egocentric and field capture when you don't have data yet
Labeling done for youOur team + AI pipeline annotate; you review and receive datasets
AI-assisted annotation (one-click smart polygon)One click finds the exact object boundary - on our own GPUs
Boxes, polygons, segmentation, keypoints, classification
Video → sampled frames for image labeling
Automatic annotation of wearable / egocentric footage7-model GPU pipeline: vision-language model, detection, pixel-accurate masks, speech, hand pose
CV exports: COCO JSON · YOLO · Pascal VOC · PNG masks
Robot-learning exports: LeRobot · RLDS · Open X-Embodiment
Model training - labeled data to deployable model weightsTrained on our GPUs from your labels or your own dataset; retrain or fine-tune with more data anytime
Dataset versioning with train/valid/test splits
Human review & curation included in the price
Your data on dedicated private infrastructurePrivate, isolated infrastructure; never pooled, never used to train third-party models

Real Human Work,Robot-Ready Datasets

Everything, Handled

Capture & Understand

Wearable Capture

AI Annotation Pipeline

Rich, Structured Labels

Label & Verify

Annotation Workspace

Human Review & Curation

Episode Intelligence

Deliver & Train

Versioned Dataset Delivery

Model Training

Private by Design

How It Works

Data collection

Annotation

Delivery

From Camera to Robot, In One Take

What Labeled Data Looks Like

Powering the Future of AI

Foundation Model Training

Robotics & Manipulation

Agent Training

Enterprise SOPs

One Service, Humans at the Core

Frequently Asked Questions