What is the AI Inspection Workspace?
The AI Inspection Workspace is a new feature in SKAND that lets users build, train, and run their own computer vision models directly inside the platform. Instead of manually inspecting every image for defects, assets, or features of interest, users can teach a model to do the first pass for them, and then review its results in a single, streamlined workflow.
This is especially useful for users working with large image sets who want to:
Detect recurring features like cracks, corrosion, stains, or equipment across a site.
Search for specific objects across thousands of images without having to scroll through each one.
Speed up repeat inspections by reusing a trained model on new project imagery.
Keep inspection data, annotations, and predictions all in one place.
This guide walks users through the full end-to-end workflow, broken into five stages:
Users can follow along from start to finish, or jump straight to the stage they need.
Important: Create an image group layer first!
Before using the AI Inspection Workspace, users need to have an image group layer already set up in the project. The workspace pulls images directly from these layers, so without one, there won't be anything to create a dataset from.
There are two ways to create an image group layer:
Import georeferenced images (typically imported with the 3D mesh/point cloud layers). See Import Images to SKAND.
Import non-georeferenced images. See How to Load Non-Georeferenced 2D Images in SKAND.
Once an image group layer is in place, users can move on to the steps below.
Create a Dataset, Select Images, and View Selected Coverage
Start the AI workflow by creating a dataset, picking images, and checking how well they cover the project area.
A dataset is the foundation of every AI model built in SKAND. It groups together all the images the model will learn from, so choosing the right images (and making sure they cover the right areas) is the most important part of this stage.
How to create a dataset and select images
Navigate to AI Datasets to see available image datasets for AI.
Click Create to Start creating a dataset.
Select Process to create dataset.
View images in image ribbon
Choose Images from Select Images Ribbon
Users can choose images from Select Images under the Image Ribbon tab.
Toggle the box to select the images.
Enable Image Coverage Frustum
View coverage of selected images – Toggles the frustums of the selected images on to display their coverage.
Image coverages are now shown.
Toggle off to hide image coverages.
Unselect images by toggling them off.
Choose Images from the Camera Widgets
Select image from camera widget – Alternatively, users can select images directly from the camera widgets displayed in the 3D Scene.
Images selected will be displayed under the Display coverage toggle.
Remove selected images from camera widgets.
Choose All Images from Select All Check Box
Users can also select all images currently visible in the scene using the Select All check box.
Give your dataset a name.
Click on Submit to create the dataset.
The dataset is ready. Next, it's time to teach the model what to look for.
Create Annotations for Training
Draw annotations on images to show the model exactly what to learn.
Annotations are the examples the model trains on. Each one is a labelled shape, usually a polygon, drawn around a feature of interest like a crack, stain, or piece of equipment. The more accurate and consistent the annotations, the better the model will perform.
Note: Before starting, make sure the project has an Annotation Template with the labels users plan to apply. This keeps everyone on the team annotating the same way.
How to create annotations
Start annotating to collect training data.
Click on an image to start annotating.
Click Annotation Toolbar and start annotating.
Draw annotation on the image.
Click Done to finalize the annotation.
Create a name for the annotation.
This indicates that the annotation and image are associated with a dataset.
Click Create to submit the annotation.
Submit the image to confirm it has been reviewed by the user and is ready for training.
Select another image to annotate.
View annotation in this image.
Submit the image to confirm it is ready for training.
Exit Process once user is done with annotation.
Click on card to view dataset details.
Filter for checked images.
View annotations in each image by clicking on the image row.
Pro tip: Quality beats quantity. A few hundred carefully annotated images usually outperform a few thousand rushed ones.
The annotations are saved and ready to feed into a training job.
Create a Training Job
Use the annotated dataset to train a custom AI model.
Once annotations are in, users can kick off a training job. SKAND takes care of all the heavy lifting in the background. Users just need to point it at the dataset and give it a name.
How to create a training job
Click Training under Create.
Select the annotation template to use for training.
Select the field with the correct classification for the annotations used in training.
Select the training dataset.
Choose the testing dataset for out-of-sample model evaluation.
Choose an existing model or create a new one.
Enter a new model name here.
Click Next to configure hyperparameters for training.
Submit training job after confirming all settings.
A pop-out notification will appear confirming that the training job was submitted successfully.
Note: Training can take anywhere from a few minutes to a few hours depending on the size of the dataset. Users can leave the page and come back. The job will keep running in the background.
Once training is complete, the model is ready to start making predictions on new images.
Create a Batch Prediction Job
Run the trained model across a new set of images to detect features automatically.
A batch prediction job applies the trained model to many images at once. This is where the time savings really show up. Instead of manually inspecting every image, the model does the first pass for users.
How to create a batch prediction job
Click the Batch Predict under ellipsis button (...) next to the process card to start batch predictions.
Select a model.
Select an AI training.
Select an annotation template for the predictions.
Choose the relevant dropdown field for the predictions.
Click Next to proceed.
Select images to add in the batch prediction data.
Toggle on Display Coverage to view the coverage of the selected images.
Submit batch prediction job.
Submission complete.
Pro tip: If users aren't sure what confidence threshold to use, start with the default. It can always be adjusted in the review stage without re-running the job.
Once the job finishes, users can review the results and fine-tune them.
Review Batch Predictions and Merge Polygons
Check the model's predictions, accept or reject them, and merge overlapping polygons into a single clean result.
This is the final quality-control step. The model has done the hard work of finding features across all the images. Now users just need to review its suggestions and clean up anything that needs a human touch.
How to review predictions
Click Start on the AI dataset to review predictions.
Select an image using the widget or ribbon to review.
Accept if prediction is good.
Review the predictions and accept them if they are accurate.
Review created annotation from accepting a prediction.
Annotation created from the prediction.
Click a different camera widget to view another image.
Merging Predictions
Merge prediction with overlapping annotation.
View another prediction.
Accept prediction if accurate and create the annotation.
Pro tip: Use the review stage as a chance to improve the next training run. If users notice patterns in the mistakes, they can add those examples back into the dataset and retrain the model.
With the predictions reviewed and polygons cleaned up, the results are ready to roll into the next stage of the project. Accepted predictions can feed straight into inspection reports, defect tracking, or asset registers, turning hours of manual image review into a clean, actionable dataset for the team.