--- title: "Image, 3D & Sensor Data | Collect, Label, Evaluate | YPAI" url: https://ypai.ai/image-3d-sensor-data/ description: "Image, 3D and sensor data collected, calibrated, labelled and evaluated in one chain of custody. Calibration, clock and consent ship with every frame." source: "src/copy/routes (route /image-3d-sensor-data/)" --- # Four sensors read one instant. Calibration makes them agree. > Image, 3D and sensor data collected, calibrated, labelled and evaluated in one chain of custody. Calibration, clock and consent ship with every frame. A camera, a LiDAR, a radar and an IMU record the same moment, each in its own frame and on its own clock. A label is only true once it survives the projection between them. YPAI collects, calibrates, labels and evaluates image, 3D and sensor data in one chain of custody. Intrinsics and extrinsics are checked per delivery, streams share one clock, and every record carries its consent, calibration and inspection. Built in Norway on EEA infrastructure, with identity-verified contributors. one clock · IEEE 1588 PTP One instant seen by a camera, a LiDAR, a radar and an IMU, drawn over the rig's frame ## Two sensors disagree until the mount is right. Re-projection error says by how much. An example rig looks at one pallet. The camera's box is the reference, and the LiDAR's cuboid reaches it through the mount you set. Move the mount and the re-projection error moves with it. Solve the calibration and the two agree. example rig · synthetic geometry · every number is computed in this tab ### Or read your own frame. Drop a photograph. yolov10n draws the camera's box and Depth Anything v2 reads relative depth from the same pixels, which lifts the box into a depth slab. Both models run in this tab and the file stays in your browser. This is monocular depth from one photograph: a model's first pass. depth · Depth Anything v2 · relative, monocular The detector found no object it is confident about in this frame. Try one with a vehicle, a person or an everyday object. - [IEEE 1588-2019 (PTP)](https://standards.ieee.org/ieee/1588/6825/) - [Re-projection error (OpenCV)](https://docs.opencv.org/4.x/dc/dbb/tutorial_py_calibration.html) - [IoU and AP (COCO)](https://cocodataset.org/#detection-eval) ## Six axes decide whether the data holds up in deployment. You name the range the system has to work across. The rig is built to span it, and every frame lands tagged on all six, so the coverage in a delivery can be counted. - The instruments on the bar, matched to the ones the product ships with. - Where each instrument sits and what it can see. Mount height, baseline, field of view, range. - The conditions that break a model trained on clear days. - The floor the product runs on, recorded at the rig's own height. - Whether the rig is still or moving, and how fast the world moves past it. - One frame, a sequence, or a return to the same coordinate across seasons. all six axes recorded per frame the rig's own frame, at rest every pass is computed from this frame's depth this condition needs its own recording · the frame is shown as captured rendering needs WebGL2 · the number is computed here as captured · this condition needs its own recording returns computed from a monocular depth map of this frame · fog, night, glare and the moved mount rendered on your GPU from that depth ## Six services, each with its own measure. Each frame below gets a model's first pass in your browser, in the label type that specialist delivers. The service behind it carries the guideline, the human review and the acceptance measure. ### [Image annotation](https://ypai.ai/annotation/image-annotation/) - Boxes, polygons, keypoints and attributes on the camera frame, with the guideline and the record. - IoU per box · agreement per batch ### [Semantic and instance segmentation](https://ypai.ai/annotation/semantic-segmentation-annotation-services/) - Pixel masks with boundary rules, drawn by people only for medical, aerial, 3D and microscopy. - mIoU · Dice · boundary F ### [LiDAR and 3D point cloud](https://ypai.ai/annotation/lidar-3d-point-cloud-annotation-services/) - 7-DoF and 9-DoF cuboids, per-point semantics and tracking, calibration-aware labels. ### [Sensor fusion](https://ypai.ai/annotation/sensor-fusion-annotation-services/) - One label held across camera, LiDAR, radar and thermal, on one clock with the IMU's ego-motion. ### [Medical imaging](https://ypai.ai/annotation/medical-imaging-annotation-services/) - Expert-reviewed labels under a task protocol, reader qualifications on file. ### [Robotics and industrial vision](https://ypai.ai/annotation/robotics-industrial-vision-annotation-services/) - Perception labels for robots and visual quality control on the line. ### Image collection - General, multi-view, document, industrial, environment-specific, rare-event and geospatial imagery. ### Sensor and IoT collection - Point clouds, depth, IMU, device deployment, synchronised sensor and video, wearables. ### [Geospatial capture](https://ypai.ai/geospatial-data-solutions/) - Aerial and satellite imagery with the pose and trajectory data that places it. - [Image data](https://ypai.ai/image-data/) - [Video data](https://ypai.ai/video-data/) - [Physical AI data](https://ypai.ai/physical-ai-data/) - [Speech and audio data](https://ypai.ai/speech-data/) - [The annotation operating model](https://ypai.ai/data-solutions/annotation/) instances and classes, one mask each a 7-DoF cuboid on classified returns one label on three streams, the error drawn the part's pose from its mask six services · the models run in this tab six services · read in this tab by six services · run the demo on this frame in your browser loading the model, then reading in this tab… Run the demo in your browser · three models, about 30 MB boxes, polygons, instance masks, the cuboid and the pose were read by yolov10n, EdgeTAM and SlimSAM in your browser as the frames scrolled in · each frame lifted on its own depth map · the frames stay in this tab · in a delivery this is the first pass a reviewer starts from ## Every frame ships with what was done to it. Capture, consent, calibration, label, inspection, acceptance, delivery. The record is written as the work happens and travels with the file. Hold the button to see how a reviewer accepts a frame. Illustrative example. A real delivery replaces every value. - cam0 30 Hz · lidar0 10 Hz · radar0 20 Hz · imu 200 Hz · one PTP clock - contributor verified by identity · release v3 · purpose written in - intrinsics K · extrinsics lidar0→cam0 · re-projection error 0.41 px - 1 cuboid · projected to cam0, cam1, radar0 · SAM 3 first pass · 6 of 8 vertices moved by a person - two of three reviewers agreed · agreement on the batch 0.91 · hidden control task passed - accepted by a named reviewer · rubric v4 - manifest · sha256 per file · Croissant · Article 10 data card · erasure within 30 days - [EU AI Act Article 10](https://artificialintelligenceact.eu/article/10/) - [Article 113, application dates](https://artificialintelligenceact.eu/article/113/) - [Croissant (ML Commons)](https://docs.mlcommons.org/croissant/) - [KITTI calibration files](https://www.cvlibs.net/datasets/kitti/setup.php) ## A model draws first. A person decides. Some lanes are human from the first vertex. SAM 3 reads a concept from a text prompt and draws the first pass on the server. The reviewer decides, and the file records which vertices the model drew. Medical, aerial, 3D and microscopy work is drawn by people from the start. - SAM 3 first pass, human review and adjudication - first pass where the concept is nameable, human on every defect - credentialed readers from the first vertex A network of 210,000+ contributors across 50+ countries, verified by identity. Each project recruits from that network for the sensors and scenes it needs. - recorded per contributor, per purpose, per session, withdrawable - territory, term and withdrawal handling defined - named at scoping and written into the release - EEA processing; subprocessors and transfers defined per engagement - within 30 days of contract end, with execution evidence - [GDPR Article 7](https://gdpr-info.eu/art-7-gdpr/) - [SAM 3 (Meta, arXiv 2511.16719)](https://arxiv.org/abs/2511.16719) ## What the rig can do now. ### [Concept pre-labelling](https://arxiv.org/abs/2511.16719) - SAM 3 segments every instance of a concept from a text or exemplar prompt, on the server, as the first pass. ### [4D radar and camera on one target](https://arxiv.org/abs/2601.21454) - A unified calibration and auto-labelling framework transfers camera labels to sparse radar points by geometric projection. ### [Splats as the scene](https://arxiv.org/abs/2602.10173) - 3D Gaussian Splatting renders a captured scene in real time; interactive splat selection puts a person in the loop of extracting objects from it. - arXiv 2308.04079 · ArtisanGS, arXiv 2602.10173 ### [One clock on the bar](https://standards.ieee.org/ieee/1588/6825/) - IEEE 1588 Precision Time Protocol synchronises the instruments' clocks so a frame, a sweep and an IMU tick can be placed on one timeline. ## Every industry mounts a different set of sensors. The table shows what a typical programme carries, and each row opens the industry's own page. The scoping call fixes yours. - [Automotive AD and ADAS](https://ypai.ai/solutions/automotive/) - [Robotics and AMR](https://ypai.ai/solutions/robotics/) - [Industrial inspection](https://ypai.ai/solutions/industrial-energy/) - [Medical imaging](https://ypai.ai/solutions/healthcare/medical-imaging-annotation/) - [Geospatial and aerial](https://ypai.ai/geospatial-data-solutions/) - [Document AI](https://ypai.ai/document-ai-workflow-automation/) ## Two sentences start the scoping. Name the sensors, the task and where the model fails today. An engineer or delivery lead reads the brief and replies within one EU business day, starting with what is feasible. An engineer or delivery lead who has read your brief replies within one EU business day. ### Describe the rig and the task - The sensors on the bar, what the model has to do, where it fails, and how the data will be used. ### Feasibility read - Scenes, sensors, calibration needs and volume checked against the network. You get a plain answer on what is feasible. ### Pilot delivery - Representative frames with their calibration, consent and inspection records, judged against your acceptance criteria. ### Production - Weekly deliveries with per-sensor QA, versioned manifests and the record for every frame. ## Calibration, consent and audit readiness ### What ships with an image, 3D or sensor delivery? Every frame carries its capture metadata on one clock, the consent record of the contributor, the calibration in force (intrinsics and extrinsics with the re-projection reading), the label with the provenance of each vertex, the inspection result and a SHA-256 hash. The delivery carries a versioned manifest, a Croissant description and a data card written in the order EU AI Act Article 10 asks for, and a signed data processing agreement. The sensor fusion service shows how one label is checked across streams. ### How are the sensors kept on one clock? The rig's instruments are synchronised with IEEE 1588 Precision Time Protocol where the hardware supports it, and every record carries the clock offset observed at capture. The tolerance a project needs is set at scoping as a service specification and checked on every delivery. It is written into the acceptance criteria. ### Which parts does a model draw, and which does a person? SAM 3 concept pre-labelling draws the first pass on the server from a text prompt. A reviewer decides in the native review console, and the file records which vertices the model drew and which a person moved. Medical, aerial, 3D and microscopy work is drawn by people from the first vertex. A model output is a first pass; what is delivered is the reviewer's decision. ### When do the EU AI Act data-governance obligations apply? Article 10 sets the data-governance duties for high-risk systems: documented collection processes, origin, preparation including annotation, and examination for bias and gaps. Under Article 113 as amended, the Annex III obligations apply from 2 December 2027 and the Annex I obligations, which cover AI embedded in type-approved products such as vehicles, from 2 August 2028. YPAI writes the data card in that order now. ### Where is the data processed, and what happens when a contributor withdraws? Primary operations are in Norway on EEA infrastructure, with subprocessors and transfers defined per engagement. When a contributor withdraws consent, the affected records are quarantined, the withdrawal is reported to you with a ledger entry, and erasure completes within 30 days with execution evidence. Delivery controls are described on the ethical framework page. Bring the rig, the dataset or the failure case. If YPAI is not the right fit, we will say so directly.