Every packaged video analytics product detects a fixed list of things. Safety Guard looks for helmets and vests, Traffic Monitor counts vehicles, Retail Counter counts people across a line. If the thing you need detected is a cracked weld bead, a palm fruit ripeness grade, a missing bolt on a conveyor bracket, or a pallet stacked above a height limit, none of those products will find it, and no amount of configuration will teach them to.
This article covers what happens after that answer. It describes which tasks fall outside packaged detection, what determines whether a custom class is buildable at your site, what the build actually consists of, and where custom development gives you nothing that a cheaper fix would not.
WHERE PACKAGED DETECTION STOPS
What A Fixed Detection Class Actually Means
A packaged detector ships with a trained class list. The model has been shown a large number of examples of a helmet and a large number of examples of a head with no helmet, and it has learned the boundary between them. That boundary is fixed at training time. A configuration screen lets you draw zones, set schedules, choose a confidence threshold and route alerts. None of those controls add a class.
This matters because the gap is usually invisible in a sales conversation. A buyer asks whether the system can detect “unsafe stacking” and hears yes, because the system detects people and the demo showed a zone rule. Zone rules answer the question of where something is. Detecting whether a stack is unsafe is a question of what a thing is, and that requires a class the model was trained on.
The distinction is worth pushing on early. A rule built from existing classes, such as a person entering a marked exclusion zone during a machine cycle, is configuration work and costs configuration money. A new class is a build.
THREE SHAPES OF TASK THAT FALL OUTSIDE THE CATALOGUE
The Tasks That Come Up Most Often
Custom enquiries at ARSA cluster into three recognisable shapes. The engineering effort differs sharply between them, and it is worth knowing which one you have before you ask for a quote.
A Class That Exists Nowhere Public
Palm fruit bunch ripeness, textile weave defects, specific casting porosity patterns, a company-specific uniform or badge. Public datasets contain none of this, so the model has to be trained on your footage. Data collection dominates the timeline here, because you need examples of the failure condition, and failure conditions are by definition uncommon. A site producing one defect per thousand units has to record for a long time, or stage examples deliberately, before there is anything to train on.
A Common Class Under Uncommon Conditions
Person detection in a boiler house at 60 degrees with steam. Vehicle counting through monsoon rain on an unlit access road. Forklift detection in a warehouse lit to 30 lux. The class is ordinary and the conditions destroy the accuracy a public model would report. This is the cheapest of the three to fix, because a general model can usually be adapted with a few thousand frames from your own site rather than trained from zero.
A Judgement That Combines Several Observations
Whether a lockout tagout procedure was followed, whether two people were present for a lift, whether a spill was cleaned within the required window. There is no single object to detect. The work is a sequence of detections plus temporal logic plus a definition of the rule that survives contact with real footage. Most of the difficulty here is specification, and it is the category where a feasibility assessment most often ends in a recommendation to narrow the scope.
THE PHYSICS COMES FIRST
Your Camera Decides What Is Possible Before Any Model Does
A custom class inherits every physical constraint in ARSA-CVS-001, the camera and site requirements standard. Pixel density governs everything: if the camera places too few pixels across the object at the distance where detection must happen, the information is absent from the image, and training cannot recover it. Pixel density falls as one over distance, so doubling the standoff halves it.
The published figures for known classes give you calibration anchors. Helmet presence needs 262 pixels per metre; hi-vis vest presence needs 153. A custom class sits somewhere on that scale according to how fine the distinguishing feature is. A 20 mm crack needs considerably more density than a missing pallet, and that number gets established during assessment rather than assumed.
| Requirement | Figure | Why It Binds A Custom Class |
|---|---|---|
| Sensor resolution | 2 MP minimum, 4 MP recommended | The minimum is the floor ARSA will contract accuracy against |
| Frame rate, main stream | 15 fps minimum | Below this, fast-moving targets are missed between frames |
| Light at target, attribute tasks | 50 lux minimum | Colour and texture cues collapse below this level |
| Camera tilt below horizontal | 10 to 25 degrees optimal | Steeper angles hide the surface carrying the feature |
| Tilt ceiling, person attributes | 35 degrees for helmets, 40 for vests | Above 45 degrees the task is prohibited outright |
| Accepted reference geometry | 3.5 m mount, 6.0 m standoff, 22 degree tilt | The worked case that meets 262 PPM at the far gate line |
| Outdoor enclosure | IP66, IK08 recommended | Applies to any custom deployment outdoors |
Two consequences follow that buyers frequently learn late. Infrared illumination produces a monochrome image, so any custom class that depends on colour stops working at night unless white light reaches the target plane. And a camera mounted high for tamper protection has to sit further back from the area it watches, which costs pixel density at exactly the point where you need it. The lens and field of view visualiser will tell you what a given lens covers at a given distance before anyone visits the site.
WHAT THE BUILD CONSISTS OF
Four Things That Decide Whether It Works
Training is the smallest part. The four things that determine the outcome are data, labelling discipline, the integration path, and ownership of accuracy after handover.
Data means footage from your site under the conditions the system will face, including the bad ones. A dataset recorded in daylight in the dry season produces a model that degrades in the wet season, and the degradation shows up as missed detections that nobody attributes to seasonality for months.
Labelling discipline means one written definition of the class that every annotator applies the same way. Where three people disagree about whether a marginal example counts, the model learns the disagreement and the accuracy figure becomes unstable.
The integration path means deciding before the build where a detection goes. A result that lands in a dashboard nobody watches has the same operational value as no detection at all. Detections route into the VMS, access control, ERP or alerting that operators already use, and that routing is a design input at the start.
Ownership of accuracy means someone measures drift and retrains against it on a contracted cadence. Gartner has predicted that 60% of AI projects lacking AI-ready data will be abandoned through 2026, and the pattern behind that number is usually a pilot that worked once and was never maintained.
Models are trained using standard deep learning frameworks and compiled for the hardware you already run. Where your existing compute is insufficient, that gets said in the assessment.
WHERE CUSTOM DEVELOPMENT IS THE WRONG ANSWER
Three Cases Worth Refusing
Custom development is the wrong answer when a packaged product already covers the task and the buyer has been describing it in unusual words. Helmet and vest compliance, vehicle counting, footfall and queue measurement are all solved by the AI Box Series with a configuration exercise.
It is also the wrong answer when the cameras cannot see the feature. A 4 mm lens at 14 metres yields roughly 96 pixels per metre against a 262 PPM requirement, and the fix is a different lens, a different position, or a second camera. Model work cannot substitute for any of those.
And it is the wrong answer when the task is performed once a week by one person in five minutes. A custom detector has a build cost, a maintenance cadence and an operational footprint. Payback depends on how much of the task is currently manual, at what labour cost, and how many sites the same model serves once it exists.
HOW AN ENGAGEMENT IS STRUCTURED
The Cheap Steps Come First For A Reason
A Remote Camera Design Review costs $1,500 and takes one week. It returns a marked-up camera schedule with a pass or fail per position against ARSA-CVS-001 and a priced remediation for anything failing. An On-Site Camera Survey costs $4,500 plus travel and covers up to 50 cameras with the same output, surveyed in person.
A feasibility assessment costs $4,500 over two weeks and produces an operational diagnosis, a technical feasibility judgement against your actual site conditions, a cost model and a go or no-go recommendation. An assessment that concludes the project should not proceed is a successful assessment: you get the reasoning in writing and you have spent $4,500 instead of a project budget. Where you contract within 90 days, the fee is deducted from the project fee.
From there, a pilot runs eight weeks on a bounded scope measured against agreed KPIs, and a production programme follows at 12 weeks or more. Each stage carries forward, so nothing gets repeated or rebuilt. The full stage table sits on the services overview.
Frequently Asked Questions
How Much Of Our Own Footage Do You Need?
It depends on how rare the condition is and how variable your site is. Adapting an existing class to difficult conditions often needs a few thousand frames. A class that exists nowhere public needs enough examples of the positive condition to cover the range of appearances it takes, which for a rare defect can mean weeks of recording. Volume, labelling responsibility and confidentiality terms are all agreed at the feasibility stage.
Who Owns A Model Trained On Our Data?
That is agreed in writing before work starts. Your data remains yours in every case. A model trained exclusively on your data for your use case is typically yours to use, while the base architecture and pretrained weights remain ARSA’s.
Can It Run On The Cameras And Servers We Already Have?
Usually. Target hardware is a design input from the beginning. Cameras below the 2 MP and 15 fps minimum can still be connected and will still produce output, with performance stated on a best-effort basis rather than contracted. Where your existing compute is insufficient, that is said in the assessment.
Does The Video Leave Our Network?
No. Processing happens on site, and video leaves your premises only where you explicitly configure it to. What crosses the network is analytic metadata: counts, events, and where enabled, small cropped images.
What Accuracy Figure Will We Get, And What Does It Attach To?
An accuracy figure is set at the pilot stage against a measured baseline on your footage, because a number quoted before anyone has seen your conditions is guesswork. It attaches to detection zones recorded as ACCEPTED in the Site Acceptance Record. If a camera is later moved or lighting is altered, those zones revert to unaccepted status until re-surveyed.
How Long From First Conversation To Something Running?
Two weeks for the assessment, then typically eight weeks to a pilot running on a bounded scope with measurable results. Data collection can extend the front of that where the condition you need detected is rare.
Bring Us The Task That Did Not Fit
If your detection problem has been quoted as impossible, or quoted as trivial by someone who never asked about your camera angles, the assessment is the cheapest way to find out which it is. Start at Custom Computer Vision Development, or contact us with the task, your camera positions and one representative clip.