A computer vision quote almost always prices the visible part of the work. The items that make a budget drift are the ones nobody mentions at the moment of signing.
This article sets out that question. It extends the article on vision and the GDPR.
What the budget really covers
Five items compose the expenditure of a computer vision project.
The scoping of the need as well as the definition of the categories.
The collection or else the acquisition of the necessary images.
The description of those images by trained people.
The modelling work as well as the evaluation of the system.
And the complete integration into the final operating environment.
One important practical consequence follows. The third item is often the heaviest share, while attention goes spontaneously to the fourth.
What the description cost depends on
Five factors determine the price of a described corpus in computer vision.
The task retained as well as the geometry it imposes.
The average number of objects really present per image.
The real difficulty of settling the field’s limiting cases.
The level of expertise required of all the people involved.
And the intensity of the quality control effectively applied.
One important observation follows for a computer vision project. Those five factors multiply rather than add, one same volume of images costing ten times more depending on the choices made at scoping.
What computation cost represents
Five expenses compose a project’s computing share.
The initial training of the model on the corpus finally built.
The successive trials run during the tuning phase.
The storage of the images as well as the successive versions.
The permanent execution of the system on the real flow.
And the periodic retraining conducted after commissioning.
One important observation follows for a computer vision project. The fourth expense ends up exceeding all the others, a training costing a few days of computation while a continuous execution consumes resources every hour for years.
What quotes most often omit
Five items disappear from the initial estimates of a computer vision project.
The preparation of the data well before any description.
Drafting the reference as well as training the teams.
The rework following the various quality controls.
The adjudication time asked of the client on ambiguous cases.
And the maintenance of the system well after commissioning.
One practical consequence follows. The last item appears in almost no project budget, while it continues as long as the system stays in use.
What the billing unit changes
Five pricing modes occur among providers.
A price per image processed, entirely independent of the content.
A price per object described, directly sensitive to density.
An hourly rate, wholly independent of the output obtained.
A fixed price covering the whole batch ordered.
And a monthly subscription for a reserved capacity.
One observation follows. Those five modes make two quotes incomparable without reworking, the unit retained counting more than the unit price displayed.
What rework costs
Five causes trigger a rework of production in computer vision.
An ambiguous reference discovered after several batches produced.
A description convention modified mid-project.
A control revealing an inconsistency between operators.
A whole category forgotten at the moment of scoping.
And an output format incompatible with the system receiving it.
One important practical consequence follows for a computer vision project. The first cause produces the most expensive rework, an ambiguity left unsettled at the start affecting every batch already produced.
What the negotiated volume permits
Five effects accompany an increased order.
The scoping cost then spreads across more units.
Training the teams is amortised over the whole duration.
The operators’ output improves markedly with habit.
The tools are configured once for the whole batch.
And the unit rate falls appreciably as a result.
One important practical consequence follows for a computer vision project. Those five effects explain why a small trial batch costs proportionally far more, which does not make it any less indispensable before committing to a large volume.
What quality control costs separately
Five arrangements are priced in addition to the production.
A systematic re-reading of a clearly defined sample.
A complete re-examination of a batch to measure omission.
A crossed double description on the difficult cases alone.
An adjudication by a referent on the disagreements observed.
And a complete correction of the batches rejected by the control.
One important observation follows for a computer vision project. The second arrangement stays the only one revealing what is missing, and is nonetheless removed first when a budget must be cut, which leaves the corpus’s most expensive defect in place.
Three errors proper to costing
Three computer vision defects belong to the budget itself.
Two quotes compared without checking their billing unit.
An amount retained with no quality control appearing in it.
And a maintenance absent from any budget forecast.
Those three defects are committed when choosing a provider, they reveal themselves several months after signing, and their common point is to compare different scopes while believing one is comparing prices.
What the client must specify to be quoted
Five pieces of information condition a reliable computer vision quote.
The task expected as well as the output format required.
A genuinely representative sample of the images to handle.
The average number of objects present per unit of data.
The level of quality control wanted on the production.
And the total volume to handle as well as the calendar envisaged.
One important practical consequence follows for a computer vision project. The second element is worth more than a long written description, fifty real images permitting a costing ten pages of specification would not make more reliable.
What this chapter shares with the previous ones
Three findings recur from the opening of this course.
Scoping decides far more than the technical execution.
Describing the data weighs more than the modelling.
And an ambiguous reference affects the whole corpus produced.
One important observation follows for a computer vision project. Those three findings translate into amounts here, which makes visible in a budget what the preceding chapters described in terms of method.
What this subject has that is particular
Four traits distinguish this question from the course’s others.
It is decided by comparing documents rather than systems.
It bears on scopes far more than on prices.
It rewards the cheapest bid over the most complete one.
And it is corrected badly once the order is committed.
One important practical consequence follows for a computer vision project. The third trait explains many disappointments, a quote excluding quality control looking more attractive than one including it while describing a markedly inferior service.
What the trial batch must establish
Five measurements come out of a reduced production before committing.
The real time effectively devoted to a unit of data.
The proportion of cases the reference does not settle.
The gap between two people describing the same images.
The number of questions sent back to the client during the batch.
And the rework rate observed after the first control.
One important observation follows for a computer vision project. Those five measurements cost a few days and make the whole budget reliable, a trial batch revealing difficulties that would otherwise have been discovered after weeks of production.
What bringing it in-house changes
Five charges appear when an internal team takes over the production.
The recruitment as well as the training of the people assigned.
The description tooling as well as its daily administration.
The daily supervision of all the production.
Keeping the activity going through the troughs in workload.
And the loss of competence occurring when people leave.
One important practical consequence follows for a computer vision project. The fourth charge often decides the arbitration, an internal team staying to be paid between two projects while an external service is sized to the real need of the moment.
What the calendar does to the budget
Five deadline constraints make a production more expensive.
A short deadline requiring far more people to be mobilised.
An accelerated training markedly reducing the initial output.
A compressed quality control mechanically raising the rework.
An insufficient client availability blocking all the adjudications.
And batches delivered far too fast to be usefully corrected.
One important observation follows for a computer vision project. The second constraint often cancels the gain sought, a team doubled but trained in haste producing less than expected while costing the whole reinforcement.
What maintenance represents over time
Five expenses continue after commissioning.
The periodic description of a recent sample meant to measure.
The supplement to the corpus bearing on newly encountered cases.
The successive retrainings as well as their validation.
The tracking of performance as well as the handling of alerts.
And the corrective interventions conducted on the installed equipment.
One important observation follows for a computer vision project. Those five expenses accumulate each year and end up exceeding the initial cost, which makes a budget stopped at delivery far below the arrangement’s real cost.
What return on investment presupposes
Five elements permit a cost to be set against a gain.
The human time the system actually comes to replace.
The cost of the errors the arrangement permits avoiding.
The volume handled set against the service’s earlier capacity.
The share of cases still being handled manually.
And the period over which the system will stay genuinely relevant.
One important practical consequence follows for a computer vision project. The fourth element sharply reduces the expected gain, a system automatically handling most cases leaving a human chain in place for the rest, which rarely removes the post concerned.
What comparing two quotes requires
Five reworkings make two proposals comparable.
Bringing all the amounts to one same billing unit.
Verifying the task described is identical on both sides.
Reintegrating quality control where it appears on a separate line.
Adding the rework expected at the rate each one announces.
And estimating the internal time each option will require.
One important observation follows for a computer vision project. The last reworking sometimes reverses the ranking, a cheaper service requiring twice as much internal support finally costing more than the proposal set aside.
What an insufficient budget produces
Five arbitrations occur when the envelope shrinks.
Quality control disappears well before any other item.
The corpus then narrows to the most frequent situations alone.
The reference is drafted faster and far less precisely.
The evaluation set is built from the existing corpus.
And maintenance leaves the scope retained entirely.
One important practical consequence follows for a computer vision project. Those five arbitrations produce an apparently compliant and durably fragile system, each one removing precisely what would have revealed the problem before commissioning.
What spreading over time permits
Five splits stagger the expenditure of a computer vision project.
A reduced trial batch produced before any volume commitment.
A first tranche covering the main categories alone.
An intermediate evaluation deciding the project’s continuation.
A targeted supplement bearing on the weaknesses observed.
And a distinct provision for the annual maintenance.
One important observation follows for a computer vision project. The third split is the most useful decision point, an evaluation conducted midway permitting a project to be stopped where the data will not reach the threshold aimed at.
What the client spends without counting it
Five internal charges escape project budgets.
The scoping time effectively devoted by the business teams.
The adjudications asked for on the ambiguous cases met.
The re-reading of all the deliverables before their validation.
The change support conducted with the users.
And the daily follow-up of the system once it is in service.
One important observation follows for a computer vision project. Those five charges appear on no invoice and nonetheless weigh heavily, a project looking inexpensive as long as the days consumed internally are not counted.
What the data provider brings here
Four practices make a costing usable.
A measurement of the real time per unit on a trial batch.
A quote distinguishing production, quality control and rework.
A comparative costing of two candidate tasks on the same need.
And a flag when the order exceeds the need expressed.
One important practical consequence follows. The first practice replaces any theoretical estimate, a time measured on fifty real images being worth more than a catalogue rate applied to an assumed volume.
What this chapter presupposes of the provider
Five practices distinguish an honest quote from a loss leader.
Measuring the real time on a trial batch before costing.
Detailing production, quality control and rework separately.
Announcing the items the quote deliberately does not cover.
Flagging that a less rich task would meet the need expressed.
And refusing a fixed price as long as no reference is written.
One important practical consequence follows. The third practice raises the apparent amount and avoids later conflict, an item announced as excluded never becoming the subject of a disagreement at the moment of invoicing.
Approaching the costing of a project
Five questions scope the costing of a computer vision project.
Exactly what task. It decides the unit cost.
How many objects per image. Density changes everything.
What expertise required. It multiplies the rate.
What quality control. It is costed separately.
And what maintenance afterwards. It is missing everywhere.
Those five answers determine a costing’s reliability. Asking them before comparing quotes avoids choosing the cheapest bid on a different scope.
The question that frames the budget
One question determines the order of magnitude of a computer vision project.
Must the corpus be created or does it already exist somewhere.
To be created, a collection to organise then a full description, commits a budget dominated by human work and a calendar of several months.
Already available, usable archives or a close public corpus, reduces the expense to verification, supplement and evaluation.
That question is settled by examining the existing data, it is asked before any costing, and it separates two projects whose budgets differ by an order of magnitude.
Three decisions before costing
Three decisions commit the budget of a computer vision project.
Verifying that a simpler task would not meet the need.
Having a trial batch produced before any volume commitment.
And requiring a written reference before the first order.
Those three decisions cost a few days, they precede any production, and their absence produces a budget whose gap with reality is discovered too late to correct.
Three checks on a quote received
Three controls qualify a computer vision costing.
The billing unit retained and exactly what it covers.
The explicit presence of quality control within the amount.
And the fate reserved for any rework and corrections.
Those three controls are each asked in one question, they require no technical competence, and their absence makes any serious comparison between two proposals impossible.
Three questions before a quote
Three questions precede any comparison of proposals.
What exactly does the billed unit cover within this amount.
Is quality control included or billed separately.
And who bears the cost of a rework after control.
Those three questions are asked in a few minutes, they require no technical competence, and their absence leads to retaining the cheapest proposal on the narrowest scope.
What this chapter teaches
One cross-cutting observation deserves closing this examination.
The billing unit counts more than the amount displayed.
Three findings compose it.
Describing the data often weighs more than the modelling.
The omitted items concentrate before and after the production.
And an ambiguous reference costs the whole corpus already produced.
That finding extends the preceding chapters, scoping deciding the budget before any execution.
What this chapter leaves to the next
A properly costed project can fail for other reasons.
Three questions stay open in computer vision.
What makes a project fail despite sufficient means.
At what moment those defects most often appear.
And which signals permit spotting them in time.
Those three questions belong to the causes of failure, which constitute the subject of the following chapter.
Why the incomplete quote wins
One point about how these decisions get made belongs at the close.
The proposal that leaves things out usually looks better.
Three reasons follow in computer vision.
An excluded item lowers the total without changing the headline task.
Whoever compares the quotes rarely knows what should be in them.
And the missing work reappears as a cost nobody attributes to the choice.
One important practical consequence follows for a computer vision provider. Listing what the quote does not cover is worth losing some bids, since the omitted work happens anyway, and a client who discovers it mid-project reads it as a provider adding charges rather than as a scope they never bought.
What each side has to bring to the costing
Four inputs make an estimate mean something.
The client supplies a representative sample of real images.
The provider measures the real time on a trial batch.
The client states the quality level the use requires.
And the provider prices control and rework as separate lines.
One important observation follows for a computer vision project. The first input has no substitute, a provider quoting from a written description alone guessing at density and difficulty, which is precisely where the figure turns out wrong.
Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.
- Comparing two quotes without ever checking the billing unit.
- Costing the modelling alone while neglecting the data description.
- Omitting the data preparation from the budget initially retained.
- Forgetting the maintenance cost after commissioning entirely.
- Choosing a task richer than the need requires.
- Neglecting the adjudication time asked of the business teams.
- Providing no margin at all for rework following the control.
- Ignoring object density when estimating the volume to handle.
- Underestimating the cost of a particularly rare business expertise.
- Signing a fixed price with no written reference beforehand.
What to take away
The billing unit counts more than the amount displayed.
Three readings emerge. Describing the data often weighs more heavily than the modelling, which moves budget attention to an item technical teams barely look at. The omitted items concentrate before and after the production, preparation and maintenance escaping almost every initial estimate. And an ambiguous reference costs the whole corpus already produced, which justifies investing a few days in a document nobody bills for.
For the causes of failure, the article on projects that fail details it. For the tools available, the article on open source in vision sets it out.
To explore delivery arrangements, supported formats and applicable control mechanisms, see our dedicated page on 2D and 3D annotation. And if you are preparing a vision project, let us discuss your need.
