Plain-language explainers on the data decisions behind better models, written by the team that ships the data.
First-person video, recorded from the wearer's own point of view, is the closest thing to human perception for an AI. This guide explains what it is, how consent-driven collection works, and which systems benefit most: ADAS, robotics, assistants, and workflow AI. Includes the capture protocol basics and what to ask a vendor before buying.
Every serious data buyer should run a pilot before committing a budget. This guide walks through the four steps we use on every engagement, send a sample, get it labeled, review against your own acceptance criteria, then scale with a fixed-scope quote, and the warning signs that a vendor won't let you test their work first.
Models trained on data without documented consent carry legal and reputational risk. This guide explains what consent-first collection actually means operationally, explicit agreements captured per record, auditable evidence, and privacy-aware handling, and why it makes your model more defensible, not just more ethical.
Bounding boxes, polygons, segmentation, keypoints, classification, temporal labels, if you're buying training data, you need to speak the vocabulary. This guide defines the core annotation types, when each is the right tool, and what to check in a vendor's labeling schema before you commit.
License-plate recognition is deceptively simple, until it meets real-world light, weather, speed, and occlusion. This guide breaks down plate detection versus character recognition, the annotation requirements (plate boxes, character-level labels, regional formats), and the quality controls that decide whether your model works in the field. Backed by the LPR work we've delivered.
Egocentric video, also called first-person or POV video, is footage recorded from the camera wearer's own point of view. Instead of watching a scene from the outside, the model sees exactly what a person sees, in the order they see it, including natural head movement, gaze, and hands in frame.
That perspective matters. A self-driving car, a warehouse robot, a kitchen assistant, or an AR wearable all operate in the same physical world a person navigates. Training them on third-person footage teaches them to perceive a scene; training on egocentric footage teaches them to act within it, because the data contains the person's actions, choices, and context, not just the environment.
When do you need it? If your system must understand what a person is doing and what they're likely to do next, task recognition, action grounding, human-robot collaboration, workflow AI, safety monitoring, egocentric data is usually the right signal. If you only need static scene understanding, standard image and video datasets may suffice.
What to check before you buy: how consent is documented per record (essential for first-person data), how capture guidelines standardize camera setup and task coverage, whether contributors span diverse real environments, and what QA criteria reject unusable footage before it reaches your team.
Read about our egocentric video collection pipeline and how we capture real-world task data.
"High accuracy" is the easiest claim in the data industry and the hardest to verify. The fastest way to cut through it is a pilot: a small, representative batch labeled against your own guidelines, reviewed against your own acceptance criteria, before any large commitment.
The four steps we use on every engagement: (1) you send a small sample of your data plus your labeling guidelines; (2) a domain-matched team annotates it with full quality control; (3) you review the output against your criteria, labels, format, consistency, turnaround; (4) if it passes, you get a fixed-scope quote for the full dataset with no hidden rework.
Warning signs that a vendor won't stand behind their work: they won't label a test batch before a contract, they have no documented quality-control process, they can't show you real samples (anonymized under NDA is normal, nothing at all is not), or they treat your guidelines as a suggestion rather than the spec.
Every engagement starts with a pilot, so you can review real, anonymized outputs from your own guidelines before you buy.
Data provenance is becoming a compliance issue, not just an ethics talking point. Regulators, enterprise buyers, and model-evaluation teams increasingly ask one question: can you prove how this data was collected?
Consent-first collection means consent is an operational artifact, not a checkbox. Concretely: contributors agree to data use before recording; the agreement is captured and retained per record; identifiable information is handled under the project's confidentiality rules; and the collection protocol follows your guidelines with documented QA rejection criteria.
Why it makes your model more defensible: if a dataset's provenance is ever questioned, in procurement review, an audit, or a regulatory inquiry, documented consent turns an anecdote into evidence. It also protects you from the higher-risk alternative: quietly sourcing data whose origins can't be verified.
We treat consent as part of every collection engagement, see our security and compliance page for the full controls list.
Annotation turns raw data into training signal. The labels you choose determine what your model can learn, so the vocabulary matters at contract time. Here are the core types and when each is the right tool.
Bounding boxes, rectangles around objects, used for detection. Cheap and fast, but coarse: a box includes whatever is inside it. Polygons, tighter object outlines for irregular shapes like vehicles or cells. Semantic segmentation, a class label for every pixel, for tasks like land-cover mapping or organ boundaries. Instance segmentation, the same, but distinguishing each individual object. Keypoints, labeled points (joints, landmarks, corners) for pose and structure tasks. Classification / tagging, whole-image or per-region labels. Temporal labels, time-stamped events for video: action starts, object tracks, behavior segments.
What to check in a vendor's schema before you commit: whether the label taxonomy matches your own classes and attributes; whether the format (boxes, segments, timestamps, metadata) fits your training pipeline without conversion; whether edge cases are documented in the annotation guide; and whether batch-level quality reporting lets you audit what you received.
License-plate recognition looks simple, find the plate, read the characters, until it meets real roads: glare, rain, dust, motion blur, skewed angles, and dozens of regional plate formats.
Two separate problems: detection (locating the plate in the frame, usually a bounding box) and recognition (reading the characters, often character-level boxes plus the string). Training data needs both: plate-level boxes, character-level labels, and ground-truth strings, ideally across day/night, weather, and camera angles.
Why quality control decides the outcome: a single misread character can send a charge or a fine to the wrong person. That's why our LPR work applies multi-tier review with explicit rejection criteria and rigorous quality controls for how we audit every batch, ensuring high-quality outputs.
Xyntriq