AI you can check

AI you can check

How Omniris finds objects in your drone photos and maps with AI, and how a person checks and corrects every result before anyone relies on it.

Summary

Drone flights produce thousands of photos and maps far larger than anyone can study inch by inch. Omniris uses AI to find the things you care about in them: vehicles, buildings, solar panels, machinery, debris, livestock, or anything else you can name in a few words. Each object found comes back outlined, boxed, labelled and counted.

AI is fast, but it is not always right. So Omniris treats every AI result as a first draft. A person can check it, correct it and mark it Verified. The check is recorded, with who made it, when, and exactly what they changed. The AI's own answer is never overwritten, so both statements survive: what the AI said, and what a person confirmed.

Why AI results need a person

Object detection, the kind of AI that finds and outlines things in a picture, has become remarkably capable. It still makes two kinds of mistake. It misses things that are there, and it marks things that are not. Seen from above, a shadow can pass for a vehicle, a skylight for a solar panel, a tarp for a pool.

For a count on a dashboard, a few mistakes may not matter. For an insurance claim, a compliance report, a dispute or a work order, they do. The usual answers both fall short:

• Trust the AI. Fast, but nobody can say which results were checked, and errors travel into reports unseen.
• Check everything by hand, outside the system. Accurate, but the corrections live in someone's spreadsheet, and the record no longer shows what the AI found or who changed it.

Omniris does both inside one record. The AI does the first pass. A person checks the result in an editor built for it. Their corrections become a separate, verified version, and the image's custody history records the check. From then on the verified version is what everyone sees: in the image viewer, on shared links, in exports and in comparisons between dates.

The detection models

All three models are open-vocabulary: they are not limited to a fixed list of objects. You describe what to find in words, and the model looks for it. Each model also traces an outline around every object it finds, not just a box, so you can see exactly which pixels it means. You pick the model at the top of the AI Detection dialog; Grounded SAM 2 is selected unless you choose another.

Choosing one

• Start with the default for everyday objects such as vehicles, buildings, people and trees.
• Try SAM 3 for less common aerial subjects, or when you want a Category preset for a particular trade.
• Choose OWLv2 + SAM 2.1 when your organization needs a model whose licence places no restrictions on use.

You can run more than one model on the same images. Each image keeps one result per model, side by side, and the image viewer lets you switch between them. Running a model again on images that already have that model's results replaces them; the dialog warns you first, and says which other models' results will be kept.

Telling the AI what to look for

The prompt is the list of things to look for, typed into Objects to detect. Separate the classes with full stops (commas work too):

The default prompt. Your last prompt is remembered in your browser for next time.

Capitals and repeats are ignored, so "Car. car." is one class. The classes you name become the labels on the results, and later the list a person picks from when correcting them.

Writing a good prompt

• Use plain object names of one to three words: "solar panel", "shipping container", "dump truck". A description such as "the panel nearest the ridge" does not work.
• Name only what you expect to see. This matters most for Grounded SAM 2.
• Condition words are the weakest kind of prompt. "Flooded road" asks the model to judge a state, not just find a thing, and results need closer checking.
• On SAM 3, every concept is priced. Under the prompt, the dialog shows the sum: concepts times images, and the credits that come to at the chosen tile size. From eight concepts up the line turns amber and, if the job costs more than the standard rate, says that removing concepts you do not need lowers the price. When a job is already at the standard rate, it says fewer concepts will not make it cheaper.

Categories

For SAM 3 and for OWLv2 + SAM 2.1, a Category menu above the prompt offers ready-made prompts for common kinds of work. Picking one replaces the prompt with that Category's list; you can still edit it, and the menu then reads Custom. Each Category is deliberately short, because on SAM 3 every noun adds work. Where a Category has limits you should know about, a note appears under the prompt when you choose it.

Grounded SAM 2 has no Category menu. Nothing is blocked: you can type any of these words into its prompt yourself.

Running a detection job

Start a job from wherever the images are:

• A folder: AI Detection on the folder's menu runs every image in it.
• Chosen files: Run AI on one image's menu, or select several images and press Run AI on the selection bar (up to 5,000 files at a time; for more, run the folder).
• Straight after an upload: tick Also run AI object detection in the upload wizard.
• From Home: the AI Detection card asks for a folder.

The dialog opens at once and fills in the image count and price as it works them out. Nothing is charged until you press Start.

flowchart LR
  A[Photos, folders
or orthomosaics] --> B[Model, prompt,
tile size, thresholds]
  B --> C[Price shown
before you start]
  C --> D[Each image cut into
overlapping tiles]
  D --> E[Outlines, boxes,
labels, confidences]
  E --> F[A person checks
and verifies]
  F --> G[Findings, change,
exports, reports]
  

What can be run

• Photos in JPEG, PNG or TIFF.
• Orthomosaics and other map images in GeoTIFF or Cloud Optimized GeoTIFF (an orthomosaic is one large, map-accurate image stitched from many photos). These are read a tile at a time at full resolution, so even a very large map does not need to fit in memory. Images stored as 16-bit or multi-band data are converted to ordinary colour for the model.
• Subfolders are included only if you tick Include subfolders; the dialog shows how many more images that adds. The output folders of map processing (stitching) are left out of a subfolder run, so a folder's own maps and models are not analysed by accident.
• Encrypted images are decrypted for the job only; the job card shows the progress as "Decrypting".

Tiles

A detector looks at a picture of a fixed size. A large photo or map is therefore cut into overlapping squares, called tiles, and each tile is examined at full detail. This is what lets a small object in a big image be found. By default neighbouring tiles overlap by a fifth.

Thresholds

• Box threshold: the lowest confidence, from 0 to 1, at which a detection is kept. Higher gives fewer, surer detections; lower gives more, with more false alarms. The default is 0.45, or 0.2 for OWLv2 + SAM 2.1. Once you type your own value, changing model leaves it alone.
• Text threshold: how sure the model must be that an object matches one of your words, rather than just looking like an object. Default 0.25. Not used by OWLv2 + SAM 2.1.

Images that were already checked

Images whose results a person has verified are skipped by default, and not charged for. The dialog says how many were skipped. Tick Include verified images to run them again: their verified version stays what everyone sees, the new AI result is kept beside it, and the image is marked Changed since verification so someone can review it. Only people who could verify in that folder can tick it.

While it runs

The job appears on the AI tab with a progress bar and a plain status, such as "Queued — waiting for GPU". You can cancel it there, view or download its log, and retry a failed job with the same model, prompt and settings. The job carries on if you close the page, and you get an email when it finishes or fails. Results then appear on each image.

What comes back

Every object the AI finds comes back with four things:

• an outline: the exact pixels the model believes belong to the object (a segmentation mask);
• a box around it;
• a label: which of your words it matched;
• a confidence: the model's own score for how sure it is, shown as a percentage. It is a score, not a probability.

In the gallery, each image's card shows how many objects were found, with a Verified badge once a person has checked it. Open an image and the AI Detection panel shows the rest:

On a large orthomosaic the outlines are fetched in pieces as you pan and zoom, so the whole map's outlines never have to load at once.

The job card on the AI tab gives the same results for the whole job: Open the folder, Export detections for every image in the job, and Findings for what the job's automatic findings created.

Orion, the Omniris AI assistant, can answer questions about the results in plain language: which images contain a class, how many objects were found, how many images are verified, or what one model found that another missed. See Orion, the AI assistant.

Checking and correcting

Verification means a person has looked at an image's AI result, corrected what was wrong, and confirmed it. It happens in the image viewer, in the same panel as the results.

Who can verify

Anyone who can edit the image's folder, including a client working on results a pilot delivered to them, in a vault they can edit. People with view-only access see the verified version and who verified it, but cannot change it. Editing and verifying are free.

The editor

Press Edit. The image becomes yours to edit for as long as you work on it, and anyone else who opens it sees that you are editing it. If you left it open on another device, you are asked whether to continue here.

The label list is the words the AI job was asked to look for, plus any class the AI returned. People correct results in the job's own vocabulary, so a "truck" never becomes a "lorry" on the next image.

The first time anyone edits an image, its AI outlines are turned into editable shapes ("Preparing outlines…"). Very detailed outlines are simplified to at most 5,000 points. On a large orthomosaic this is done in the background; if it cannot be done, you can still correct the boxes.

A draft, then Verified

• Your edits are a draft. Each change is saved within a second, and the draft is still there when you come back. Nobody else sees it as the result until you verify. Discard draft throws it away.
• Verify publishes your corrected version in one step. Verify & next does the same and opens the next image in the folder that still needs checking, in the gallery's order, skipping any that someone else is editing. When none are left, it says so.
• The AI's own result is never changed. The panel switches between Verified (what everyone now sees) and AI original (what the AI said), so the two can always be compared.
• Verifying again starts a new draft from the verified version. The next check becomes verification 2, then 3, and every earlier one stays in the image's history.

The custody record of each check

Each verification is written as a sealed event in the image's tamper-evident custody record, the same record that tracks its upload, processing and sharing. The event records:

• who verified it and when, and which verification of the image this is;
• which AI job's result was checked;
• how many objects were added, removed, moved, resized, relabelled, and had their outline changed;
• digital fingerprints of the result before and after the check, so either can be proved unchanged later.

The panel's History lists every verification with its change summary, and links to the custody record. Custody certificates for the image list the AI result and each verification. The companion paper Evidence you can prove explains the custody record and certificates.

What a verification records, in summary. Illustrative example values.

Working through a folder

• Badges on each image card: Verified; Changed when something new arrived after it was verified; Replaced since verification when the picture itself was replaced.
• The Verification filter in the gallery toolbar: All images, Not verified, Verified, or Changed since verification.
• The folder count under the file list, for example "212 of 480 verified · 12 changed since verification".

Verified versus changed

A verification stays in force until something real changes underneath it:

• Changed since verification. A newer AI result for the image, or an updated delivery from a pilot, arrived after it was verified. The verified version still stands and is still what everyone sees. In the editor, the new AI result is one more view, and objects it found that the verified version lacks are offered as New in the AI re-run, each to accept or leave.
• Replaced since verification. The image's pixels were replaced. A check of the old picture says nothing about the new one, so the image goes back to Not verified. The earlier verification stays in the record.
• Verified by someone else meanwhile. If another person verified the image while you were editing it, the editor stops and offers a review of both versions, keeping yours or theirs.

Detections that are already findings

When a detection you change is the evidence for a finding, the editor names the finding and asks what to do: Keep the finding, or mark it a False positive. With Keep, the finding follows your correction: its evidence picture is re-cut from the corrected box and it is moved to the corrected place, unless someone set its location by hand. A deleted detection stays in the finding's evidence, marked as removed in that verification. Each decision is applied on its own, so one that cannot be applied (for example, a finding under legal hold) is skipped and listed, and the verification still goes through.

Copies, deliveries and shared links

• Delivered and copied images carry the AI result and its verification, shown as "Verified by [name] on the delivered original" or "on the original". From then on each side verifies its own copy.
• Shared links show the verified version with "Verified" and the verifier's name, never an email address.
• Downloaded images carry their detections, whether they are verified, the verification number and its date in the image's embedded metadata (XMP). The verifier's name is not included; it stays in the custody record.
• Removing AI results from a verified image asks first, naming who verified it and when. The AI's raw result is removed and the verified result stays. Images under a legal hold or retention policy keep their results.

Checking in the field

Checking is often best done close to the site, by the person who flew it. The Omniris phone app and desktop application use the same editor as the web, so the tools, the rules and the record are identical wherever the check is made.

• On the desktop application: the same AI Detection dialog, editor, Categories, findings choices and change comparison as the web.
• On the phone: start an AI job on a folder (model, prompt, tile size, thresholds and Include verified images, with the price shown first), open an image's results, and correct and verify them full screen. The gallery's Verification filter and Verify & next work there too.

Verifying with no signal

The phone's editor is built into the app, so it opens without a connection. To prepare a folder for a site with no coverage, choose Make available offline for verifying on the folder while you still have a connection. The app downloads each image and its AI results, with a permission to verify it offline that lasts 30 days.

Some images can only be checked on the phone while online:

• Orthomosaics can be corrected and verified on the phone, but only with a connection.
• Encrypted images are never stored on a phone, so they are not offered offline.
• Images without an AI result, and images you cannot edit, are not offered either.

Signing out deletes the downloaded images and results from the phone. Your unsent drafts are kept, hidden from anyone else, and resume when you sign back in. The companion paper the phone app in the field covers the app as a whole.

AI change between dates

When the same site has been mapped on two dates, Omniris can compare what the AI found on each and tell you, object by object, what appeared and what disappeared. It works in the Compare view of the Site Navigator: pick the two dates and a map layer, choose the model, set the confidence, and press Start.

The results are drawn on the map: appeared in green, disappeared in red, and unchanged in grey if you choose to show them. Click an object to see its confidence on each date, how far it moved and its margin of error. The summary gives the counts and the shift between the dates. Distances are in feet, with metres in brackets.

Each comparison is recorded as an event in the later image's custody record. Objects that appeared or disappeared are filed as findings awaiting review, and the comparison can be printed in the situation report (see Reports and exports). A comparison reuses stored results, so it is not charged for.

From detections to findings

A detection says "there is a tarp here". A finding is the issue someone has to act on: placed on the map, with a category, a severity, evidence pictures, an owner and a status. Omniris turns detections into findings in several ways:

Detections close to each other are grouped into one finding, and ones near an open finding are added to it rather than duplicating it. The choice When the job finishes is remembered for the organization or vault; When each image is verified must be chosen each time.

The companion paper Findings you can act on describes findings in full: triage, assignment, evidence, history and reports.

Reports and exports

Exporting detections

Export detections is on an image's AI panel, on a folder (with every folder below it), and on an AI job's card. It writes the detections as GeoJSON, CSV or KML, for mapping (GIS) software, spreadsheets and Google Earth.

• The checked version. Where a person verified an image, the export holds the verified detections; otherwise the AI's. A verified column says which, detection by detection.
• On the map. Each detection is given a location: exactly, on an orthomosaic; on a photo, projected onto the ground from where the camera was and where it pointed, or at the camera's position when its direction is not recorded. Each comes with a margin of error, in feet and metres.
• Nothing silently dropped. A detection that cannot be placed is still exported in GeoJSON and CSV, with its box in image pixels and the reason. KML can only hold placed objects, so it states how many were left out.
• Delivery. Download it from a link kept for 24 hours (1 hour when any source image is encrypted, because the export itself is not), or save it into a folder you choose, where it gets its own custody record. Exports are free.

The situation report

Generated from the Compare view for a site and two dates, the situation report is one PDF that brings the comparison together:

• the area on the later date, with an outline, scale bar and north arrow;
• both dates side by side at the same scale, the change in ground height where both dates have a survey, and the AI change counts;
• each changed object, with its picture from both dates, and the other open findings in the area, those nobody has reviewed first;
• the sources: capture dates, custody, content fingerprints, and a custody certificate for each map that can be checked offline;
• a Method and limits section: the model and its licence, the classes looked for, the confidence cut-off, what "unchanged", "appeared" and "disappeared" mean, and a plain statement that AI detections can be missed or wrong and that unreviewed items have not been checked by a person.

It is written in English or Polish, and saved as a new file in the later map's folder (or your own Reports folder) with its own custody record. Encrypted imagery is never printed into it.

Elsewhere

• Custody certificates list the AI result and every verification of an image, with who started the job and who verified it.
• Downloaded images carry their detections and verification status in embedded metadata (see Checking and correcting).

For sharing folders, links and delivery to clients in general, see sharing, delivery and integrations.

Price

AI detection is priced by the number of images, in blocks of up to 100. The standard rate is 14 credits per block, rounded up to whole blocks: 1 to 100 images cost 14 credits, 101 to 200 cost 28, and so on. The dialog shows the price, and your credit balance, before you start.

• What you see is what you pay. The price in the dialog and the amount charged come from the same calculation. If your balance is too low, the job is refused with the amount needed, and nothing is charged.
• Verified images skipped by a run are not charged for.
• Failure costs nothing. A job that fails or is cancelled is refunded in full.
• Checking is free. Editing, verifying, exporting detections and comparing two dates carry no charge.
• Hybrid. On your own hardware, a job sent to the Omniris cloud is billed to your organization's Omniris cloud credits.

Where it runs

Detection runs on a processing machine with a graphics processor (GPU), not in your browser. Once started, the job is carried through by Omniris itself, so you can close the page.

A site running Omniris on its own hardware can also connect to the Omniris cloud. The dialog then offers Run in Omniris Cloud (Hybrid), job by job: the images for that job are processed in the Omniris cloud and the results come back to the site. It is never chosen for you.

Editing and verifying need no GPU and work in every deployment, on the web, the desktop application and the phone.

What the results claim, and what they don't

AI detection is a fast, consistent first pass. These are its limits, stated plainly.

• It misses things and marks things that are not there. It is a detection aid, not a census. Check against a known image before relying on counts, and verify what matters.
• Confidence is the model's own score, not a probability. Scores differ between models: those of OWLv2 + SAM 2.1 run lower than the others.
• Wording matters. Plain object names work best. Conditions ("flooded", "damaged") are the weakest kind of prompt, and thin linear objects such as power lines come back in pieces.
• It finds objects, not their condition. The Solar PV Category finds panels, not cracks, soiling or hot spots; the roof Category finds roof features, not damage.
• Big objects can split. An object wider than a tile can come back as several pieces. On SAM 3, choose a larger tile size for roofs, roads, fields and water.
• Running a model again replaces that model's earlier AI result on those images; the dialog warns first. Verified versions are kept, and other models' results are untouched.
• Size limits for checking. An image with more than 10,000 objects, an outline of more than 5,000 points, or a single save over 20 MB opens read-only in the editor; its results still show. A draft that reaches its size limit asks you to verify before making more changes. An unverified draft nobody has touched for 30 days is deleted.
• Locations are estimates. On an orthomosaic a detection's place is exact to the map; on a photo it is projected from the camera and carries a margin of error, which is always stated.
• Change between dates needs two orthomosaics with the same model and prompt, at ground resolutions no more than twice apart. It compares what was detected, so anything the AI missed on either date is missed in the comparison too.
• On the phone, orthomosaics are checked online only, encrypted images are never stored offline, and the Category menu and the automatic-findings choice are not offered when starting a job; use the web or desktop application for those.
• The situation report is a plain PDF. It is not made for a pair whose later map, or its folder, is encrypted.
• SAM 3's licence asks you to acknowledge SAM if you publish results based on it, and prohibits military and unlawful use. The two other models have no such conditions.

Where it matters

Glossary

Box threshold

The lowest confidence at which a detection is kept.

Category

A ready-made prompt for a kind of work, such as Solar PV or Construction site.

Changed since verification

A verified image for which a newer AI result or an updated delivery has arrived; it is still verified, but due for another look.

Concept

One thing named in a prompt, such as "dump truck". SAM 3 looks for each concept separately.

Confidence

The model's own score, from 0 to 1, for how sure it is about a detection. Not a probability.

Custody record

The tamper-evident history of a file: who uploaded, processed, shared, checked or changed it, and when.

Detection

One object the AI found: an outline, a box, a label and a confidence.

False positive

A detection of something that is not there, or is not what the label says.

Finding

An issue to act on, placed on the map, with a category, severity, evidence and status.

GeoTIFF

An image file that also records exactly where on Earth it lies.

Open-vocabulary

Able to look for anything you describe in words, not just a fixed list of objects.

Orthomosaic

One large, map-accurate image stitched from many overlapping drone photos.

Outline (segmentation mask)

The exact pixels the AI believes belong to an object, rather than just a box around it.

Prompt

The list of things you ask the AI to look for.

Text threshold

How sure the model must be that an object matches one of your words.

Tile

One of the overlapping squares a large image is cut into, so each part can be examined at full detail.

Verification

A person's check of an image's AI result, sealed in its custody record. The AI's own result is kept beside it.