Model Garden

Production-ready computer vision models for utility infrastructure inspection, plus your team's adaptive models, pipelines, and BYOM detectors.

No Adaptive Models

Create an adaptive classifier in Model Builder and it will appear here for testing and API access.

Create Adaptive Model

No Pipelines

Build a pipeline from production and adaptive models, then test it here.

Open Pipeline Builder

No BYOM Models

Enterprise users can upload custom detection models and use them in tests and pipelines.

Go to Enterprise Dashboard

AI Studio

Model Builder
Model Garden
Pipeline Builder
MCP Integration
API Docs

Get Started

About Adaptive AI
Pricing
Enquire
FAQs

Legal

Terms
Privacy Policy
Cookie Policy
Information And Data Security

Model

Output Labels

VT quota: -/-

Prompts Threshold

API Access

Authenticate requests with an API key. Create an API key to call this model from your application.

Select a model to see API details

Options

Crop margins

Expand each detection's crop bounding box beyond its default size. Affects bounding boxes only — an instance outline is never moved, so a mask stays on the shape the model traced.

Crop Background

Applies to the crop thumbnails and to what Save as dataset writes. Saved crops never carry the highlight outline — that is for viewing only.

Remove background Background colour

Options

Mask threshold

Pixels at or above this confidence count as wire. Lower catches fainter wires, higher cuts false positives. Also applies to results already back, without re-running them.

0.50

Max input size

Longest side sent to the model, in pixels. Lower is faster — cost scales with area — but below about 2048 thin conductors start to fragment. Below 1280 costs no less than 1280, since the model pads to one tile.

4096

Source

Drop images here or click to browse

Run batch inference on local images

Limit to images

Results

No results yet

Choose a model, add images, and run inference to see results here.

Save Results as Dataset

What to save

Full images containing detections

Cropped detections (as new images)

Limit

max images

Dataset name

Saving...

Model performance evaluation

Upload a ground truth JSON file to measure model metrics such as precision, recall, F1, and more. The ground truth file can be uploaded before, during, or after inference.

Couldn't load this model's output labels. Close and reopen this dialog to try again.

Upload

Drop a ground truth JSON file here, or click to choose one

Flexible matching (ignore folder path, file extension, and capitalization)

View JSON Example

Example

{
  "IMG_0001.jpg": "person",
  "IMG_0002.jpg": "cat",
  "IMG_0003.jpg": "dog",
  "IMG_0004.jpg": "tree",
  "IMG_0005.jpg": "car"
}

Model metrics show up once both the ground truth file and inference results are ready.

Ground truth is ready. Add images or select a dataset to start evaluating them against it.

Coverage

0 images missing ground truth

If an image is being reused from a previous upload under a different filename (content-based dedup), make sure the ground truth key matches the filename shown in this batch, not the original upload's name.

0 ground truth keys with no matching image

0 ground truth entries with an invalid label

0 ground truth keys collide under flexible matching

These keys normalize to the same filename (same name ignoring path, extension, and capitalization), so there's no reliable way to tell which one is correct. None of them are used, and any image matching this filename is excluded from metrics. Rename or remove the duplicates to resolve this, or turn off flexible matching.

0 images share the same filename after normalization

These images normalize to the same filename as each other (same name ignoring path, extension, and capitalization), so there's no reliable way to tell which one a ground truth entry for that name actually describes. None of them are used, even if the ground truth file has an entry for this name. Rename the images to make them distinct, or turn off flexible matching.

Metrics

Model metrics will start showing once you run inference.

Exclude classes with 0 ground truth and 0 predictions

Macro average metrics Weighted

Per class metrics

Confusion matrix

Confidence calibration

Confidence threshold trade-off

Advanced filter

Filter by model predictions

Each group is a set of conditions that all must be true at once (AND). A crop or image matches if it satisfies any one of your groups (OR). Multiple groups let you express alternative combinations, such as requiring two conditions together in one group, or a single different condition in another.

With no groups at all, every image/crop is included. Thresholds set in Additional filters below always apply, whether or not you've added any groups here.

Additional filters

Filter by name