Detecting & selecting objects
Detection is the step that turns a scene image into a list of separate objects — and the step where you decide what the scene is going to cost.
Running detection
In a Bulk project, Detect Objects is already wired to the Scene Generation node. Connect your scene image to it and run it.
You can optionally give it categories — the kinds of object to look for. Leave it empty and it will find whatever it can; supply categories and it will focus on those. Categories are worth using when a scene has one class of object you actually care about, for example furniture in a room full of props.
Detection returns a set of objects with a preview. Nothing else runs yet, and nothing else is charged yet.
Reviewing the selection
Open the object viewer to see what was found. Everything is selected by default, so this step is about removing what you don't want.
Every object you keep is charged for all three generation stages that follow, so the selection screen is the cheapest place to control your spend. Work through the list and deselect:
Structure — walls, floors, ceilings, ground planes. These are usually better built in your DCC than generated.
Anything small or distant. 3D generation needs each object to fill most of its reference image. An object that was tiny or heavily occluded in the scene won't produce a usable reference, and will usually fail at the 3D stage.
Backdrop — anything blurred or only partly visible that you wouldn't place in a level.
Duplicates — if the scene contains six near-identical crates, keep one and instance it yourself.
Fragments — a detection that's caught part of an object rather than the whole thing.
Confirm the selection when you're happy with it. The count you confirm is the count you're billed for at each downstream stage.
Running detection again
If the detection isn't right — it missed objects you wanted, split one object into several, or merged several into one — you can run Detect Objects again. Each run is charged the same as the first.
This is worth doing before you generate. Re-running detection is cheap relative to what follows, so fixing a bad decomposition costs far less than generating an object that was never a whole object in the first place.
Objects missing entirely
Name them explicitly in categories
One object split into pieces
Re-run with a broader category for that object
Several objects merged into one
Re-run with the individual categories named
Too much scenery detected
Just deselect the extras — that's free
If two runs don't improve things, the scene image is usually the problem rather than the detection. A clearer image with better-separated objects, each one reasonably large in frame, will detect better than any category list. See Sourcing & preparing assets.
What happens next
Once you confirm a selection, the generation stages run per object, as a group:
Generate Reference Images
Generate 3D Model
Generate PBR Textures
Each stage runs across every object in the selection. See Working with scene results for what to do when some objects succeed and others don't.
Last updated