AI and data · Computer vision
Computer vision development services for video, images and inspection
We build computer vision systems that watch video, find and track objects, read text in images and flag defects, then put the result in front of the people who act on it. We choose between custom detection models, vision-language models and cloud vision APIs per task, and decide early whether inference runs on the device or in the cloud.
Short answer
Computer vision development means building software that understands images and video: detecting and tracking objects, reading text, inspecting products and reviewing footage. Innovation Insight builds these on YOLO-class models, vision-language models such as Gemini, GPT and Claude, or cloud APIs like Rekognition and Vertex AI. Discovery costs $3k to $8k, a computer vision MVP $25k to $60k, and a full platform with devices, dashboards and reporting $50k to $150k+.
Reviewed by Zain Khalid Malik, CTO & Co-founder · Updated
What we build
Computer vision projects rarely fail on the model. They fail on the footage: cameras at odd angles, night-time glare, a defect that appears twice a month, or an upload pipeline that cannot keep up with an hour of video per guard per shift. We scope around the cameras, the volume and the decision a person makes with the output.
| System | What it does | Typical approach |
|---|---|---|
| Video analytics and review | Turns long recordings into searchable events, frames, transcripts and short summaries | Frame sampling, sprite maps, a vision-language model reviewing selected frames, transcription for audio |
| Object detection and tracking | Counts and follows people, vehicles, parcels or equipment across frames | A YOLO-class detector with a tracker such as ByteTrack, on the edge or a GPU server |
| OCR in images and video | Reads labels, meters, serial numbers, plates and documents photographed on a phone | Cloud OCR or an on-device model, then validation rules |
| Visual inspection | Flags scratches, missing parts, wrong labels or misaligned packaging | A custom detector or anomaly model trained on your defects, on a fixed camera |
| Image classification and moderation | Sorts uploads, flags unsafe content, tags products | Cloud APIs or a fine-tuned classifier |
| Live monitoring dashboards | Shows the latest frame per camera, positions on a map and alerts | Event streams from devices, a web dashboard, alert rules |
Three ways to build it: custom models, vision-language models or cloud APIs
The choice depends on how specific the task is, how fast the answer must come back and how many images you process. Many systems combine two: a cheap detector finds the moments that matter and a vision-language model explains them.
| Approach | Strengths | Weaknesses | Best for |
|---|---|---|---|
| YOLO-class detectors (custom trained) | Fast enough for real-time video on modest hardware, run on the edge, cheap per frame at volume | Need labelled data for your classes; Ultralytics YOLO code and trained models are AGPL-3.0 unless you buy an Enterprise licence | Counting, tracking, inspection, anything high-volume and well defined |
| Vision-language models (Gemini, GPT-4o class, Claude) | No training data needed, describe scenes in words, answer open questions about an image | Slower and dearer per image, can describe things that are not there, not suited to every-frame analysis | Incident review, summaries, rare or varied situations, prototypes |
| Cloud vision APIs (Amazon Rekognition, Vertex AI Vision, Google Cloud Vision) | Working in days, pay per image or minute, no models to host | Generic labels, limited control, data leaves your device | Moderation, generic labels, text detection, early validation |
For reference, Amazon Rekognition's image APIs list at $0.001 per image for the first million images a month, and stored video label detection at $0.10 per minute in US East (checked October 2026). At thousands of camera-hours a month, per-minute pricing usually pushes the high-volume path towards a self-hosted detector.
Edge or cloud inference
| Factor | Edge (camera, phone, NVIDIA Jetson-class box) | Cloud (GPU server or managed API) |
|---|---|---|
| Latency | Milliseconds, works offline | Network round trip, depends on upload bandwidth |
| Bandwidth | Send events and thumbnails, not raw video | Upload everything you want analysed |
| Privacy | Raw footage can stay on site | Footage leaves the site; needs agreements and retention rules |
| Model size | Small, quantised models only | Large models, including vision-language models |
| Fleet management | Provisioning, certificates, updates and monitoring for every device | One deployment to manage |
| Typical use | Real-time alerts, inspection lines, poor connectivity | Post-event review, search, reporting, heavy models |
Body-worn and vehicle cameras are a common middle ground: devices record and upload, the cloud does the heavy analysis, and the fleet still needs proper device management with certificate rotation.
How a video pipeline works
- Upload: devices push segments to object storage; an event router classifies each upload by type.
- Fan out: separate jobs stitch video, extract audio, generate frames and sprite maps and run transcription, each with an idempotency key and a retry policy so a retried job never doubles the work.
- Detect: a detector or sampled-frame model marks the moments worth looking at.
- Review: a vision-language model looks at the selected frames and transcript and returns structured fields: severity, detections, scene conditions, key phrases and recommended actions.
- Store beside, never over: the AI review is stored next to the human record, with the provider, model, prompt and temperature, so it can be audited and reproduced.
- Report: dashboards, search and PDF reports for the people who act on the output.
Data labelling and accuracy
- Class definitions first: a written guide with examples of what counts as a person, a defect or a vehicle, and what does not. Most labelling disagreements are definition problems.
- Labelling tools such as CVAT or Label Studio, with a second reviewer on a sample of every batch.
- Hold-out sets split by site and camera, not by random frame, so the test measures performance on cameras the model has not seen.
- Metrics per class: precision, recall and mAP for detectors, plus false alarms per camera-hour, which is the number operators feel.
- Hard-case mining: every false alarm and miss from production goes back into the next labelling batch.
Mistakes we see in computer vision projects
| Mistake | What happens | What we do instead |
|---|---|---|
| Testing on frames from the training cameras | Accuracy looks excellent in the report and drops on the first new site | Hold out whole sites and cameras, and test at night and in rain as well as on clear days |
| Analysing every frame of every video | GPU or API bills scale with recording hours, most of which show nothing | Sample frames, use motion or a cheap detector to pick moments, and send only those to heavy models |
| Letting AI output replace the human record | One wrong review overwrites what the officer or inspector actually saw | Store the AI result beside the human record, with the model and prompt that produced it |
| Ignoring the upload path | Field devices on mobile data fall behind and footage arrives hours late | Chunked uploads with resume, idempotent processing jobs and per-device health monitoring |
| Measuring accuracy but not false alarms | Operators start ignoring alerts within weeks | Agree a false-alarm budget per camera-hour and tune thresholds to it |
| Leaving licensing until launch | An AGPL model in a closed product forces a rewrite or a late licence purchase | Choose the detector and its licence in discovery |
Who uses the system
| User | What they need from it |
|---|---|
| Operators and reviewers | A queue of flagged events with the clip, frames, transcript and AI notes on one screen, and quick accept or correct actions |
| Supervisors and managers | Live dashboards per site, trends, and reports they can send to clients without editing |
| Field staff and devices | Simple capture, reliable upload on poor connections, and clear device status |
| Administrators | Device provisioning, user roles, retention settings, model and prompt versions, and audit logs |
Privacy and biometrics law
Cameras capture people, and some vision features process biometric data. The rules below shape what we build and where. This is a summary, not legal advice; your counsel should confirm it for your use case.
| Rule | What it covers | What it means for the build |
|---|---|---|
| Illinois Biometric Information Privacy Act (BIPA) | Collection of face geometry and other biometric identifiers from people in Illinois: requires a written policy, a retention schedule and informed written consent | Liquidated damages of $1,000 per negligent and $5,000 per intentional or reckless violation; a 2024 amendment limits recovery to one per person for the same identifier and collection method. Avoid face templates unless consent is designed in |
| GDPR Article 9 | Biometric data processed to uniquely identify a person is special category data in the EU and UK | Needs an Article 9 condition such as explicit consent, plus a data protection impact assessment. Detecting that a person is present is not the same as identifying who they are |
| EU AI Act Article 5 | Prohibits untargeted scraping of facial images to build recognition databases, emotion recognition in workplaces and schools (except medical or safety), biometric categorisation by sensitive traits, and real-time remote biometric identification in public spaces for law enforcement except in narrow cases | Applies from 2 February 2025. We do not build these uses, and we design detection features so they cannot be repurposed into them without a new review |
Our default is detection without identification: count and track people as anonymous boxes, blur faces in stored frames where they are not needed, and keep retention short.
What computer vision development costs
| Engagement | Cost | Timeline |
|---|---|---|
| Discovery: footage review, approach and licence choice, accuracy target, fixed estimate | $3k to $8k | 1 to 2 weeks |
| Computer vision MVP: one pipeline, one model, dashboard or review screen | $25k to $60k | 10 to 14 weeks |
| Full platform: device fleet, several pipelines, reporting, roles | $50k to $150k+ | 4 to 8 months |
| Ongoing support, retraining and fleet operations | $1.5k to $8k per month | Monthly |
GPU hosting, cloud API fees and labelling hours come on top and depend on volume. The computer vision development cost guide breaks these down, and computer vision use cases shows where the approach pays off. See pricing for all published rates.
Computer vision work we have shipped
- AI video analysis for a US security services company: body-camera footage from patrolling officers is routed by an S3 event router to jobs for video stitching, audio extraction, frame and sprite-map generation and transcription, so an hour of patrol footage becomes reviewable in seconds. Google Vertex AI reviews every incident and returns proposed severity, detections, scene conditions, key phrases and recommended actions, stored beside the officer's record and never overwriting it. The platform includes live dashboards with the latest frame from each camera and GPS positions, device fleet management with AWS IoT provisioning and certificate rotation, and 1,855 automated test cases in the worker service.
- Edge OCR: AI OCR on uploaded documents and images at volume, with AWS Lambda processing, AWS Bedrock for document understanding and queue-based retries, the same text-in-images problem that appears in label and meter reading.
We also build AutoViz, a 3D and AR wheel configurator. That is real-time rendering, not computer vision, so we do not count it as vision work. If your images are mostly documents, AI document processing is the better fit; for forecasting and scoring from tabular data, see machine learning development. The wider practice is at AI and data services.
Sources and further reading
Next step
Tell us what you're building and get a written estimate.
A senior engineer replies within one business day. NDA on request.
Products we've shipped, and what happened next.
Case studies written from the technical documentation of each project: the stack, the scale and the outcome.
Questions we get asked a lot.
How much does computer vision development cost?
Discovery costs $3k to $8k. A computer vision MVP with one pipeline and a review screen costs $25k to $60k, and a full platform with devices, several pipelines and reporting $50k to $150k+. GPU hosting, API fees and labelling are extra.
Should we train a custom model or use a vision-language model?
Use a vision-language model when the task is varied, low volume or needs a written explanation. Train a YOLO-class detector when the classes are well defined and you process a lot of frames, where it is far faster and cheaper per frame.
Can computer vision run on the device instead of the cloud?
Yes, with small quantised models on phones, cameras or Jetson-class boxes. It cuts latency and bandwidth and keeps raw footage on site, but you take on device provisioning, updates and monitoring.
How much labelled data do we need?
For a custom detector, usually a few hundred to a few thousand labelled images per class, drawn from your own cameras. Vision-language models need none to start, which is why we often prototype with them.
Is facial recognition legal?
It depends on where. Illinois BIPA requires written consent and carries statutory damages, GDPR treats identifying biometrics as special category data, and the EU AI Act bans several uses outright. Our default is detection without identification.
How do you measure accuracy?
Precision, recall and mAP per class on a hold-out set from cameras the model has not seen, plus false alarms per camera-hour in production. We agree the target in discovery.