A vision SDK for the physical world: factory lines, delivery yards, perimeters, control rooms. Detection, tracking, and visual search running on hardware right next to your cameras. Millisecond latency. No cloud round-trips. No per-frame fees.
Built for operations, not for demos. Each capability runs at camera frame rate on modest hardware and composes with the others.
Find every item of interest in every frame. Ships with general-purpose models. Drop in your own for products, packaging, vehicles, or anything specific to your line.
A milk bottle on the line missing its cap. A label that did not print. A damaged pallet at the dock.
Keep a consistent ID on every moving object across frames and, with a little wiring, across cameras. Vehicles, forklifts, shipments, people in PPE.
A forklift entering a no-go zone. A worker who has not left a restricted area in ten minutes. A truck moving across eight cameras in a yard.
Compute an image embedding in a few milliseconds, index it in any vector store, and search your archive by example. Text queries too, in the same space.
Does this defect look like one we have flagged before? Find every frame in the last shift that resembles this incident.
The same SDK powers quality, logistics, safety, and security workloads. A few representative scenarios, drawn from real deployments.
Detection catches the missing cap, the misaligned label, the empty package, the blank expiry print. At 60fps per camera on a small industrial PC. An alert on the PLC or a reject signal on the actuator, with the frame stored for review.
Read plates at the gate. Follow a truck across every camera in a logistics yard with one persistent ID. Flag a container that did not leave on its scheduled slot. Works when the internet does not.
Tracking plus detection turns a wall of CCTV screens into a system that only speaks up when it matters. A fall, a flame, a person where they should not be. Alerts land in your control room in under a second.
Similarity search turns every stored frame into a vector you can query. A QA engineer uploads a defect photo and gets every matching frame across three weeks of video. A supervisor searches for vehicles that looked like the one in yesterday’s incident.
Your defect classifier, your domain-trained detector, your internal quality model: the SDK runs it like any built-in capability. Describe inputs and outputs in a short JSON block, call one function, get a typed result.
Four shapes are supported out of the box: image-to-image (segmentation, enhancement), image-to-class (classifier), image-to-embedding (search, similarity), image-to-detection (your own YOLO-style model).
// Register a custom cap-check detector. { "models": { "cap_check": { "model": "cap_detector.onnx", "type": "image_to_detection", "input": { "resize": 640, "normalize": "0.5" }, "output": { "labels": "bottle, cap_ok, cap_missing" } } } }
Point the SDK at a supported model and it runs. Ship open weights we include, swap in a fine-tuned version, or bring your own ONNX export in the same family.
Detection, segmentation, pose, classification, and oriented bounding boxes in one line.
Image and text embeddings in a shared space. Similarity search over an archive, plus text queries.
Full stack: text detection, recognition, angle classifier. Expiry dates, lot codes, container IDs, labels.
Multi-object tracking. Composes with any detector in the catalog.
Embeddings for matching the same subject across cameras. Yards, depots, multi-camera perimeters.
Open-weight face detection and embedding models the SDK loads and runs.
Need a vendor-supported full-stack FR system with imaged-trained weights, antispoof, and an SLA? See By invitation below.
Open-weight plate detectors paired with PaddleOCR for plate text. Works out of the box at gates, checkpoints, and yards.
Monocular depth estimation for AR, parallax, and 3D measurement.
Semantic segmentation families commonly fine-tuned in-house by industrial customers.
Promptable segmentation for operator-assisted QA and annotation bootstrapping.
Standard classifier backbones for BYO-classifier workflows.
4x super-resolution for low-res archives and incident frames.
Vehicles, drivable surface, and lane lines from one pass. Fits fleet, depot, and yard cameras.
Find titles, paragraphs, tables, and figures in scanned pages. Pairs with PaddleOCR for full document parsing.
A small number of imaged capabilities are licensed only to vetted customers, under separate NDAs and separate terms. Those capabilities are not described publicly, and the list is not long.
If your project requires something specific and sensitive that is not listed on this page, say so in a call. We will tell you directly what we can and cannot do, and on what terms.
Start the conversationA conveyor runs at one meter per second. A forklift does not pause because a data center is slow. A fire does not wait for a round-trip to a remote API. Inference that matters at the edge has to happen at the edge.
imaged runs on the industrial PC, the Jetson, the factory gateway, or the server rack you already own. A reject signal lands in under a frame. When the internet drops, your line keeps running. When you add a camera, you do not add a bill.
One SDK, five platforms, five language bindings. CoreML on Apple hardware, CUDA on Linux boxes with GPUs, CPU on everything else.
One fee per year, per deployment, per platform. No per-frame, per-camera, or per-user charge. Every capability is on, every version upgrade is included.
Trials are free for 30 days with a generated token. Paid tokens are issued after a short call. If your setup is unusual (many cameras on one box, or one token across a fleet), tell us on the call and we will price it up front.
