Knowledge · Architecture
Choosing vision architecture
After sensor geometry, teams still face platform, algorithm and dimensionality choices. Here is the decision tree for industrial Computer Vision, including robot vision and paths into 3D bin picking where depth is required.
Platform
Where does processing run?
Smart cameras embed tooling and I/O for standalone stations. PC-based vision scales with GPU, multi-camera setups, and custom algorithms. Most lines use one or the other, not both on the same task.
Algorithm
Classical rules or AI?
Not every defect needs a neural network. Classical tooling stays maintainable when features are stable. AI earns its place when variation, texture, or classification complexity exceeds rule-based limits.
Classical vision vs AI
When to stay with rules, when to train models, and how to keep validation practical.
Sensing dimension
2D, 3D, and robot vision
Many inspection tasks stay in 2D. Robot guidance and bin picking add depth when pose, occlusion, or height matter. Compare sensing approaches before locking a 3D camera SKU.
Application path
When architecture leads to 3D bin picking
3D bin picking is one demanding branch of robot vision, not the default for every depth sensor. If random pile presentation, occlusion, and 6D pose drive your spec, follow the guide on cycle time, pose estimation and feasibility. For flat pick-and-place or 2D alignment, stay in the guides above.
Need help mapping architecture to your line?
Describe platform constraints, cycle time, and whether depth is truly required. We will point you to the right guide or a feasibility conversation.