Support FAQs

It depends on how specific your requirements are and how soon you need the data. Off-the-shelf datasets are ready to deliver within 3–10 business days, cost less, and come with a free sample. That makes them the best fit for common tasks such as face recognition, liveness detection, speech recognition, or license-plate OCR. Custom collection is built to your exact requirements: demographics, geographies, languages, scenarios, capture devices, and annotation format. It suits niche use cases and edge cases that existing data doesn't cover. Many teams use both: they start training on an off-the-shelf dataset, then add custom data to fill the gaps.

It depends on the type of dataset. Off-the-shelf datasets are licensed: you get the right to use the full dataset to train and test your models under the license terms, while Unidata keeps ownership. For custom datasets, ownership, exclusivity and data retention are agreed in the contract, and IP rights can be fully transferred to you if your project requires it.

Each purchased dataset is delivered as a fixed version, current as of the delivery date. If your project needs more or newer data, such as new demographics, scenarios or a larger volume, Unidata can extend the dataset with additional collection, arranged as a separate order.

If a delivered batch falls below the agreed threshold, Unidata reworks it free of charge, analyzes the root causes (guidelines, calibration, annotator assignment), and reports corrective actions back to you. Persistent issues are escalated to the Quality Control Department and project management for team re-calibration or re-staffing.

Yes. For long-running or evolving projects, Unidata assembles a dedicated team — annotators, QA specialists, and a project manager — that works exclusively on your data, learns your domain, and scales with your volume. This fixed-capacity FTE model gives you a predictable monthly cost instead of per-unit billing.

Guideline changes are a normal part of ML development, and Unidata handles them through a structured change process: the team assesses the impact, identifies which already-annotated data is affected, and provides a cost and timeline estimate for re-annotation before proceeding. Minor clarifications are absorbed into the regular calibration loop at no extra cost.

Unidata collects custom training data for ML and AI across seven service lines: image, video, speech, audio, text, egocentric video, and robotics data. This covers real-world photos and biometrics, human-activity and driving footage, scripted and spontaneous speech across languages and accents, ambient and industrial sound, domain-specific text and prompt–response pairs, first-person footage for embodied AI, and synchronized sensor data for robots. Unidata also collects LiDAR and medical DICOM data (CT, MRI, X-ray), and can validate and annotate everything it collects within the same project.

Unidata chooses the method based on the data type, volume, and intended use of the dataset. Crowdsourcing sends tasks to distributed participants through crowd platforms such as Microworkers, targeted by geography, language, profession, and skills; Unidata sets the task, checks every submission, and handles acceptance. In-house collection covers structured offline capture, such as studio recording, field sessions, and on-site capture with specialized equipment like head-mounted capture setups for egocentric video. Unidata can also select and prepare open-source or marketplace datasets after checking their licenses, and render synthetic data for scenarios that are rare, dangerous, or unavailable in the real world. Methods can be combined in one project.

Depending on the task, Unidata works with consented participants, crowdsourcing platforms, in-house or on-site capture, client-provided materials, legally permissible public sources, and licensed or ready-made datasets. Before collection starts, the team checks source permissions, participant consent, platform terms, geographic restrictions, and the usage rights the intended application requires.

Outsourcing gives AI teams a ready-made operation for participant recruitment, capture, validation, and project management, so they don't have to build a temporary one in-house. At Unidata, a dedicated project manager runs each collection from brief to delivery. This shortens time to data, makes capacity easier to scale, and lowers operational risk, while keeping the dataset aligned with your technical specification.

There is no universal minimum. The volume you need depends on the task, the model and its pretraining, data variability, label granularity, class balance, target metrics, and edge-case coverage. Unidata recommends starting with a representative pilot, measuring model performance and learning curves, and then expanding the dataset based on error analysis.

Yes. Unidata can tailor data sources, participant profiles and geography, scenarios, capture conditions, devices and sensors, volume, class balance, metadata, annotation, quality thresholds, and delivery format to your model and its intended use. The proposed design is usually validated through a pilot before production scales up.

Yes. Depending on the use case, Unidata can combine text, images, audio, video, LiDAR/3D data, and synchronized sensor streams such as depth, IMU, and pose. For multimodal capture, the setup specification defines devices, calibration, timestamps, synchronization tolerances, metadata, and output structure before production starts.

Yes. Unidata supports multilingual and multi-region collection in 32+ languages and 50+ countries, subject to participant availability and local legal and operational constraints. The specification can define language, locale, dialect or accent, geography, demographic quotas, device and environment conditions, and reviewer qualifications. Native speakers or domain-qualified reviewers are involved where required.

Yes. Unidata organizes on-site and scheduled field collection in real-world or controlled environments, subject to location access, participant availability, safety, equipment, and legal requirements. For time-sensitive projects, feasibility, launch timing, production capacity, quality controls, and delivery milestones are confirmed during scoping.

Quality is controlled at three stages: before launch, during collection, and before delivery. Unidata validates the specification and the pilot; monitors technical parameters, task completion, metadata, and quotas during production; and runs automated checks and human review against agreed acceptance thresholds. Items that don't meet them are corrected, re-recorded, or excluded under the agreed rework rules.

Yes. Validation criteria come from the technical specification and can cover completeness, file integrity, format and metadata, technical quality, scenario or task compliance, labeling consistency, and — where relevant — sensor or stream synchronization. Checks run during collection and again before delivery, and acceptance thresholds and rework rules are agreed in advance.

Yes. Unidata's data collection is GDPR- and CCPA-compliant. Participants must explicitly confirm their participation before they can start a task, and consent forms are signed wherever the data requires it — for example, selfies, voice recordings, or any other biometric attribute. Data is stored on AWS infrastructure covered by ISO 27001 and ISO 27701 certifications, with strict access controls at every stage.

Before launch, Unidata defines the lawful basis or source permission, participant notices and consent where required, permitted uses, retention period, and transfer conditions. Personal data is minimized and can be pseudonymized or anonymized where appropriate. The exact controls depend on the data category, the jurisdictions involved, and your intended use.

Security controls are set for each project based on the data type and client requirements. Depending on the scope, they can include NDAs, role-based access, secure transfer and storage, data minimization, and agreed retention or deletion rules. When AWS infrastructure is used, data is hosted on services covered by AWS ISO 27001 and ISO 27701 certifications.

Yes. Collection, validation, and annotation can run as one workflow, so the same technical specification and metadata scheme carry through to labeling. Unidata's 1,100+ labelers and specialists work with text, image, video, audio, LiDAR/point-cloud, DICOM, and other data, with the annotation method and quality controls defined for the model task.

Unidata delivers standard and project-specific formats for text, images, video, audio, LiDAR/point clouds, DICOM, annotations, and multimodal sensor data. Before production, the file type or codec, folder structure, naming convention, metadata schema, identifiers, timestamps, and delivery method are agreed, so the dataset fits straight into your pipeline.

Timelines depend on the data type and volume, target regions, participant recruitment, site access, equipment, annotation complexity, validation requirements, and client review cycles. After a feasibility assessment, Unidata provides a project plan covering the pilot, ramp-up, expected throughput, quality review, and final delivery milestones. For reference, past projects include 2,000 biometric images in 5 weeks and 750 children's audio recordings in 1 month.

Pricing depends on the delivery unit — for example, each accepted image, audio minute, video hour, participant, route kilometer, or captured scene — and on recruitment, locations, equipment, travel, annotation, validation, usage rights, storage and delivery, and timeline. The quote can separate pilot and production costs and is tied to the agreed acceptance criteria. Every project starts with a free consultation and estimate.

The pilot is paid but priced at cost: you get measured numbers on your actual data — quality, throughput, and per-unit effort — instead of a speculative quote. When you proceed to full-scale production, the pilot cost is credited toward the contract.

Multi-stage QA is part of every engagement and is stated explicitly in the commercial proposal. Depending on the depth of validation you need — standard review, dedicated validators, or double-blind consensus — it is either included in the unit price or listed as a separate line item, so you always see exactly what level of quality control you are paying for.

All services are GDPR and CCPA compliant runing on AWS infrastructure certified under ISO 27001 and ISO 27701, with strict access controls applied throughout the annotation process.

Annotated data is delivered in the format you need — COCO, Pascal VOC, JSON, CoNLL, PCD, or a custom schema — along with a full quality report.

Unidata uses established commercial and open-source annotation platforms - e.g. CVAT, Label Studio, plus format-specific tools (LabelImg, V7, RectLabel, VoTT, Prodigy for images/video; Audacity, Sonix, Descript, Speechmatics for audio; CloudCompare, 3D Slicer, SUSTechPOINTS, Voxel51 for 3D/LiDAR) - combined with AI-powered automation and human review.

Data labeling covers basic tasks: assigning a category or tag to a whole item, such as image- or video-level classification. Data annotation is broader. It includes semantic segmentation, keypoint and landmark marking, polygon and cuboid outlines, temporal and event marking, metadata enrichment, and relationship mapping — the level of detail advanced ML models such as object detection or pose estimation need.

Image: bounding boxes, polygons, semantic/instance segmentation, keypoints, 3D cuboids, classification, landmarks, line/mask annotation. Video: object tracking, semantic/instance segmentation, action recognition, keypoint & event tracking, temporal segmentation, polylines, 3D cuboids, scene-text recognition. Audio: speech-to-text transcription, speaker diarization, sound-event detection, emotion recognition, music tagging, voice-activity detection, phoneme-level and utterance classification. 3D / Point Cloud / LiDAR: 3D bounding boxes, semantic/instance segmentation, object tracking, keypoints, mesh & volume annotation, lane/road-marking and terrain annotation.

Yes. Unidata runs projects on a Scrum-based Agile model, with iterative delivery, fast feedback loops, and flexible scaling as requirements evolve.

Timelines depend on data type, dataset size, and annotation complexity, so Unidata evaluates each project individually and provides a clear delivery schedule up front. A team of 1,000+ trained annotators, combined with AI-assisted tooling, keeps turnaround fast without compromising quality.

Each project follows a structured, milestone-based workflow: (1) kickoff briefing and task setup, (2) NDA, (3) pilot and scoping estimate, (4) tooling and workflow configuration, (5) execution by domain-matched annotators, (6) human-in-the-loop QA review, and (7) delivery of a production-ready dataset with a full quality report. Every project is supervised by a dedicated project manager.

Yes. Unidata scales human annotators, automation tools, and QA workflows to meet enterprise-level requirements across diverse datasets and 19+ industries.

Your data is handled exclusively by Unidata's managed team of 1,000+ experienced annotators with domain expertise across 19+ industries. It is never outsourced to open crowdsourcing platforms.

Every batch goes through a multi-stage QA process that combines human review with automated, AI-assisted validation. Unidata's dedicated Quality Control Department, with 6+ years of experience, reviews annotated data daily, tracks metrics such as error rate, inter-annotator agreement (IAA), and intersection over union (IoU) for spatial annotations, and benchmarks results against curated "golden" reference samples.

Unidata delivers 95%+ annotation accuracy, validated daily by the Quality Control Department. Exact accuracy targets are agreed for each project's data type and requirements before annotation begins.

Yes. Every engagement can start with a small, representative pilot batch with a clear cost estimate, so you can validate annotation quality, workflow, and compatibility with your ML pipeline before scaling to full production volume.

There is no strict minimum. Unidata supports both small pilots and large production-scale projects. Pilot batches are usually 10–100 samples, depending on task complexity; typical engagements start at 500–5,000 data points (images, video clips, audio files, etc.), and a full training dataset is commonly 5,000–50,000.

Yes. Datasets can be integrated with complementary regional or modality datasets to improve model generalization (for example, combining license-plate datasets from multiple countries for multilingual OCR).

Yes. Datasets are delivered in widely supported file formats (e.g. JPG, PNG, MP4/MOV, WAV/MP3, CSV/JSON/XML), which integrate directly with standard computer-vision and machine-learning frameworks.

After a request is submitted, the Unidata team contacts the client to confirm requirements and prepares the necessary agreement/documentation. Once the agreement is signed and payment is completed, the dataset is delivered securely (typically via cloud access) within 3–10 business days.

All datasets are securely hosted on AWS cloud infrastructure. Storage and data-management practices comply with ISO 27001 and ISO 27701 standards, ensuring high availability, scalability, and strong information-security/privacy protection.

Yes. All datasets are curated in full compliance with GDPR and other applicable data protection laws; data is collected only from legally permissible sources, and synthetic datasets contain no real personal information at all.

Yes — Unidata datasets are proprietary and collected specifically for its clients; they are not scraped from open-source repositories or otherwise publicly available.

Unidata offers both: most biometric, speech, and traffic datasets are real-world recordings from real participants or environments, while certain identity-document datasets (e.g. synthetic passports, driver's licenses) are fully AI-generated and contain no real personal data. Each product page states which type applies.

Depending on the dataset, data is collected via vetted crowdsourcing platforms, controlled in-studio recording sessions, verified data-collection partners, or by parsing real-world footage/imagery; synthetic datasets are generated with AI. In every case the sources are legally permissible and ethically sourced.

Unidata uses a dual-licensing model: a free sample is available for trial/evaluation, while access to the complete dataset is granted only after purchase.

Yes. Unidata can build a custom dataset tailored to a client's requirements — specific demographics, geographies, languages, attack/scenario types, annotation formats, volume, or recording conditions — as an alternative or supplement to the off-the-shelf datasets.

Confirm the file formats, annotation/metadata types, dataset size, and demographic or scenario diversity match your project's needs, and request the free sample first to validate quality and compatibility with your pipeline.

Yes. Unidata provides a free sample of virtually every dataset so you can evaluate data quality, file formats, and annotation/metadata structure before committing to a full purchase.