👁️ Computer Vision & Sensing
AI that sees and interprets visual and spatial information for inspection, monitoring, and analysis. Heavily clustered at leading-edge: object detection, OCR, and quality inspection have proven deployments but most organisations lack the labelled data or edge infrastructure for production scale. Only one practice reaches good-practice. Most trajectories are stalled, waiting on hardware costs and data pipeline maturity to unlock broader adoption.
The Headline
Computer vision is cheap and everywhere, but proof that it works is not. This fortnight's news was about permission: Britain promised statutory rules for police face scanning, and a Maine town banned plate readers.
The Picture
Most organizations using vision AI are buying a commodity: shelf scanning, fleet cameras, imaging triage and drone mapping are all easy to purchase from many vendors. A small group of large operators, including grocery chains, imaging networks and national screening programs, now runs these tools across entire estates. Almost everyone shares the same weakness. The outcome numbers mostly come from the vendors and operators themselves, and independent checks keep finding weaker field results, especially on rare, small or unusual cases, which are often the ones that matter. Systems that identify people face a separate test: they work well enough, so the open question is whether regulators, voters and courts will allow them, and on what terms.
This Fortnight
Britain's Home Office announced on November 4 that police facial recognition will get its own statute and an independent watchdog. Use must be intelligence-led and independently tested for bias on age, gender and race, and the watchdog could seek court injunctions to stop a force using the technology. The same announcement included £11 million to triple the fleet of live-scanning vans, which Big Brother Watch called "totally inadequate" as a safeguard. A separate code of practice will cover private businesses, so UK retailers and venues using face matching should expect rules of their own.
Voters in York, Maine passed a ban on license plate reader cameras by 73% to 27%. At least ten U.S. local governments have moved to curb or end contracts with Flock, the largest plate-reader network, since summer, and Flock has shortened its default data retention and now requires a case code for every search. If your parking lots, campuses or properties feed plate data to police, expect your retention and sharing terms to be questioned.
Spain's data protection regulator narrowed the circumstances in which biometric verification counts as necessary. The AEPD says the test is met only where no equally effective, less intrusive alternative exists, and it names qualified electronic signatures and digital ID as such alternatives. The opinion follows a €950,000 fine on Yoti earlier this year for offering no non-biometric option. In Europe, any face-based door, sign-in or attendance system now needs a non-biometric route and a written explanation of why that route would not do.
RadNet says its radiologists approve over 90% of AI-drafted reports without changes. The figure, given on its third-quarter earnings call for a draft-then-sign workflow (AI writes the report, a radiologist approves it), came with no error rate or methodology. An independent Korean review published in August found that AI-drafted chest reports carried more clinically significant errors than human-only ones. An unchanged sign-off rate measures speed, not accuracy.
Regional grocer DeCicco & Sons extended Simbe's shelf-scanning robots from a pilot to all 12 of its stores. The decision rests on Simbe's own reports of fewer out-of-stocks at the pilot stores. A new investment analysis found that no retailer anywhere has disclosed its own measured return from shelf robots. Before you sign, ask for results measured by a retailer, not by the vendor.
Coming Up
Britain's facial recognition statute is promised, not passed, and the Met plans to extend live scanning to fixed cameras in the West End and Soho in December. Live facial recognition matches passing faces against a watchlist in real time. Until the law exists, forces will keep expanding under today's patchwork of rules. When the bill appears, check whether it sets a minimum crime-seriousness threshold and what the private-sector code requires of retailers and venues.
Texas legislators return in January with bills that would ban plate readers at public agencies or require warrants to search their data. None has passed. The state already froze its Flock funding in August, and plate-reader limits are now arriving from ballots, city councils and statehouses at the same time. Map where your organization's cameras feed shared networks before a local vote or a new law decides for you.
DHS expects to award two biometric identity contracts worth $50 million to $100 million each on January 11, 2027. One pays human examiners to resolve uncertain fingerprint and face matches. The other funds independent testing of matching and anti-spoofing tools. Democrats take both chambers of Congress in January and plan to investigate DHS biometric surveillance. If the largest buyer in the market pays both for people to check the machine and for outside accuracy testing, your own procurement should ask for both.
What's Hard About This
Putting a person alongside the AI can make the person worse. In a new emergency-department study, physicians' sensitivity for dislocations fell from 80% to 18.2% when they used a fracture-detection aid, while a chest X-ray aid from the same vendor improved their reads. Each tool has to be tested on its specific task, with your own staff, before it is trusted.
Nearly every large outcome figure in this field is reported by the company that benefits from it. Kuaishou's 5.89% order lift from AI visual search came with no sample size, and RadNet's sign-off rate and Simbe's stock-out gains carry no published methodology. Budget for your own before-and-after measurement, because independent checks keep finding field results below the sales pitch.
Virtual try-on shows how clothes look, not whether they fit. A Forbes columnist found Zalando's fitting room showing a pair of jeans fastened on her body, although she could not actually wear them. A review of 315 eyewear try-on studies found no benchmark that links face scans to how frames physically fit. If size-driven returns are the problem, fix your size data before paying for better images.
Practices in this Domain (20)
Read the full technical briefing (1,589 words) →
Where AI Stands in Computer Vision & Sensing
Computer vision has become cheap, plentiful and easy to buy. Proving that it works in a given building, scanner or street is still slow and expensive. Across inspection, medicine, retail, security and mapping, deployment is now broad. RoboSense shipped 719,200 LiDAR units in the first half of 2026. The Royal College of Radiologists finds AI tools in 75% of UK clinical radiology departments. Simbe says more than 3,000 shelf-scanning robots are under contract across over 75 retail banners. Hong Kong International Airport has put more than 93 million passengers through face-verified Flight Token since 2021. What lags is independent evidence of success. Again and again, a model that scores well on its benchmark disappoints in the field. Aidoc's FDA-cleared pulmonary embolism triage reached 86.8% sensitivity across a 17-facility health system, below its clearance benchmarks. Japan's road ministry found people counting failing on about 30% of 2,427 road sections. An audit of 152 geospatial foundation-model papers found about 40% release no weights, and that published scores for the same model differ by ten points or more.
Momentum is building where the task is bounded and a human still owns the output. In radiotherapy contouring, 85% of 80 radiation oncology professionals in a self-selected survey report using AI, mostly for auto-contouring, which the authors call "broad but shallow". In mammography screening, Germany's PRAIM trial of 500,000 women reports 17.6% higher detection and 43% less radiologist workload across 40% of the country's screening centres. Shelf monitoring at large grocers is the one application that has clearly crossed into routine chain-wide use. Kroger is cited as running about 900 Tally robots, with more than 2,100 to come. Even there, an investment analysis finds no retailer-measured ROI disclosed for the robot category. Fleet camera safety and drone photogrammetry sit close behind. Their gains are real but come almost entirely from vendors and operators.
Two things set this domain apart. First, its failures cluster at the edges of the data. Sensitivity drops to 50% for brain microlesions. Wildland flame detection scores 0.29–0.36 AP50-95, against more than 0.90 for smoke in industrial fires. Face verification falls to roughly 88–92% on identical twins. Damaged 18th-century manuscripts defeated standard layout analysis entirely. Averages hide these cases, and they are often the ones that matter. Second, the applications that identify people are now limited by legitimacy rather than accuracy. Facial recognition in policing and access control, plate readers and public-space video analytics all work well enough to deploy. Regulators, voters and courts are contesting whether they should be deployed, at what scope and with what disclosure. That leaves the domain unusually split. Inspection and measurement wait on validation and data pipelines. Identification waits on consent.
What's New, 2026-11-02 to 2026-11-16
The biggest movement this fortnight was in governance, not capability. On 4 November the UK Home Office announced a statutory framework for police facial recognition. It requires intelligence-led use, independent bias testing on age, gender and race, and a new watchdog that could seek court injunctions. Alongside it came £11m to triple the fleet of live-recognition vans, which Big Brother Watch called "totally inadequate" as a safeguard. The Met now claims more than 3,000 arrests from live recognition in two years. It also reports 270,000 faces scanned at Notting Hill with one false alert, figures without a published methodology. Western Australia's concluded trial reports 122 arrests and nine incorrect alerts from almost 1.4 million faces, without demographic breakdowns. In the US, voters in York, Maine passed a ban on plate-reader cameras by 73% to 27%. At least ten local governments have moved to curb or end Flock contracts since summer, and Flock cut its default retention from 30 days to seven. Texas bills that would ban or warrant-gate public plate readers are filed but not enacted. Spain's data protection regulator, the AEPD, narrowed the necessity test for biometric verification: it applies only where no equally effective, less intrusive alternative exists. Police Scotland is running a BriefCam object-recognition pilot, with facial recognition excluded, on 100 Aberdeen cameras until 14 December. Nanjing University dropped a classroom "head-up rate" metric after leaked footage caused a backlash.
On the commercial side, the headline numbers were large and self-reported. RadNet says its draft-reporting tool now covers over 30% of its exams, and that over 90% of drafts are signed off unchanged, with no error methodology given. DeCicco & Sons extended Simbe's Tally to all 12 of its stores on vendor-reported stock-out gains. Kuaishou reports orders up 5.89% from generative visual search, without sample sizes. DHS forecast two OBIM contracts worth $50m–$100m each, due for award on 11 January 2027: one for human examiners to settle uncertain IDENT/HART matches, one to test matchers and presentation-attack tools. The counter-evidence was sharper. In an emergency-department study, BoneView assistance cut physicians' dislocation sensitivity from 80.0% to 18.2%, while ChestView lifted chest anomaly sensitivity from 49% to 79%. A Forbes hands-on test found Zalando's fitting room rendering jeans the author could not wear. A 315-report review found no benchmark linking facial scans to physical eyeglass-frame fit. A fuel-cell microscopy benchmark found five simple image statistics matched vision-language models.
Key Tensions
Self-reported scale, unaudited outcomes. The largest deployment claims this cycle came from the operators or vendors themselves: RadNet's 90% unchanged-draft rate, Kroger's robot fleet, Ailytics' 400+ sites, Kuaishou's order lift. Where independent scrutiny exists, it pulls the other way. The strongest construction-safety field result, Zurich and Arrowsight's 50% lower claim frequency, was company-reported and mixed with human video review. An investment analysis calls shelf-robot ROI a disclosure void.
Averages hide the cases that matter. Aggregate scores keep concealing failure on small, rare or out-of-domain inputs. UCSF's brain-metastasis model reached Dice 0.783 overall but 50% sensitivity for microlesions. Google Open Buildings showed 90% recall but precision of 50% or less in South Africa. A Yale study found AI matching experts on tumour volume but agreeing only at κ 0.47 on whether a tumour had grown, which is the clinical decision.
Assistance that degrades the reader. Putting a human beside the AI does not guarantee a better result than either alone. A crossover study found 71% of AI-prompted chest X-ray revisions turned correct reads into wrong ones. This cycle's emergency-department study saw dislocation sensitivity collapse with BoneView assistance. A Korean review cites 22.8% clinically significant errors in AI-drafted chest reports, against 14.0% for human-only reports, which sits awkwardly with the push towards draft-then-sign.
Legitimacy, not accuracy, gates identification. For faces, plates and crowds, the binding question is now whether use can be justified. The Home Office has promised a statute that is not yet law. The AEPD names e-signatures and digital ID as alternatives that defeat a necessity claim. FOI data showed South Wales Police selecting ethnic minority people for 24% of officer-initiated scans in 2025, against 7% of the local population. Meanwhile Britain's roughly 14,000-camera plate network draws little of the scrutiny now aimed at Flock.
Showing is not knowing. Several applications sell a visualisation as though it were a measurement. Virtual try-on renders garments that would not fit, and OpenAI and Google both disclaim fit in their own documentation. Coresight attributes over half of apparel returns to size. In geology and materials science, high predictive scores are likewise being separated from validity: one ImageNet-pretrained diffraction model reached high accuracy by reading background noise.
Top 10 Evidence Items
- Home Office to launch new watchdog and legal safeguards for police use of facial recognition (news-coverage) — Centres the fortnight's dominant movement: governance arriving in statute while capability claims stay unaudited. https://www.info-gov.uk/data-protection/data-protection-news/home-office-to-launch-new-watchdog-and-legal-safeguards-for-police-use-of-facial-recognition
- Cross-sectional study evaluating the impact of an artificial intelligence-assisted system on diagnostic accuracy for chest and bone X-rays in the emergency department (research-paper) — Direct evidence for the 'assistance that degrades the reader' tension, with a stark sensitivity collapse rather than a benchmark gain. https://link.springer.com/article/10.1186/s13049-026-01718-x
- Assessment of Vestibular Schwannoma Growth After Radiosurgery: Variability in Clinical Decision Thresholds and the Role of Artificial Intelligence (research-paper) — Shows how a headline agreement score (ICC 0.99) conceals the κ 0.47 disagreement on the actual clinical decision. https://link.springer.com/article/10.1007/s10278-026-02516-8
- Accuracy of Google’s Open Buildings Dataset in Characterising Diverse Building Configurations in South Africa (research-paper) — Classic average-hides-failure case: 90% recall sounds strong until precision at 50% or less shows the model can't be trusted for boundaries. https://www.mdpi.com/2673-7418/6/6/136
- RadNet, Inc. (RDNT) Q3 FY2026 earnings call transcript (adoption-metric) — The self-reported-scale half of the central tension: a 90% unchanged-draft rate with no error methodology behind it. https://finance.yahoo.com/quote/RDNT/earnings/RDNT-Q3-2026-earnings_call-729187.html
- Retail inventory robots technology and investment research (industry-report) — Names the ROI disclosure void directly, undercutting the shelf-monitoring deployment numbers cited elsewhere as the sector's clearest success. https://pxsresearch.com/research/physical-ai/retail-inventory-robots/
- Spain’s data protection regulator clarifies biometric risk and necessity (news-coverage) — Captures the pivot from accuracy to legitimacy: a regulator ruling biometric use unnecessary once a less intrusive alternative exists. https://www.scworld.com/brief/spains-data-protection-regulator-clarifies-biometric-risk-and-necessity
- Ethnic minority groups in Wales four times more likely to be target of police facial scans (news-coverage) — Grounds the legitimacy argument in a concrete disproportionality finding, the kind of evidence that outlasts any accuracy statistic. https://www.theguardian.com/technology/2026/nov/15/ethnic-minorities-wales-police-facial-scans-data
- AI Can Show You The Dress. But Can It Tell You If It Fits? (news-coverage) — Illustrates 'showing is not knowing': a rendered fit that visibly isn't one, with a vendor's own disclaimer behind it. https://www.forbes.com/sites/angelachan/2026/11/02/ai-can-show-you-the-dress-but-can-it-tell-you-if-it-fits/
- Scenario-Wise Comparative Evaluation of Lightweight YOLO Models for Flame and Smoke Detection (research-paper) — Documents the edge-of-data failure mode precisely: wildland flame detection scoring a third of the industrial-fire benchmark. https://www.mdpi.com/2571-6255/9/11/500