Choosing a model type

Selecting the right model type is crucial for your computer vision task. Each type is designed for specific use cases and provides different types of information.

Quick decision guide

Need to classify entire images? → Classification
Need object contours? → Segmentation or Instance Segmentation
Need to count separate objects? → Instance Segmentation
Need precise boundaries? → Segmentation / Instance Segmentation

Task Type	When to Use	Key Benefits
Instance Segmentation (Default Recommendation)	Start with this for most industrial applications	Most versatile and informative Can always simplify analysis if needed Provides geometric and spatial information More robust than classification
Segmentation	Single object or non-overlapping regions Don't need to count instances Need faster inference	Faster processing Simpler than instance segmentation Good for region identification
Classification	Truly just need image-level decision No localization needed whatsoever Very simple pass/fail scenario AND segmentation would be overkill	Fastest inference Simplest approach Minimal computational requirements

Pro Tip

When in doubt between Segmentation and Instance Segmentation, choose Instance Segmentation. It's more capable and the performance difference is usually negligible on modern hardware.

Comparison table

Feature	Classification	Segmentation	Instance Segmentation
Localization	❌	✅ (regions)	✅ (per object)
Object counting	❌	❌	✅
Multiple classes	❌	✅ Multiple regions	✅ Multiple objects
Precise contours	❌	✅	✅
Overlapping objects	N/A	❌ Cannot separate	✅ Can separate
Size measurement	❌	⚠️ Total area only	✅ Per object
Position data	❌	⚠️ Region center	✅ Per object
Speed	Fastest	Fast	Moderate
Complexity	Simple	Moderate	Complex

Real-world examples

Example 1: Bottle inspection line

Task: Inspect bottles for defects

Wrong choice: Classification → "Defective" / "Good"
Problem: Cannot tell where the defect is or how many defects exist

Better choice: Instance segmentation → Detects each scratch, dent, crack separately
Benefit: Know exactly what defects exist, where they are, and their severity

Example 2: Presence detection

Task: Check if a product is present on the line

Option A: Classification → "Present" / "Absent"
Option B: Segmentation → Segment the product region

Recommendation: Use Segmentation

• More robust (uses shape and position)
• Can also tell you if product is correctly positioned
• Provides size information
• Can be converted to simple presence/absence

Example 3: Counting pills

Task: Count pills in a tray

Wrong choice: Classification → Cannot count
Wrong choice: Segmentation → Blends touching pills together

Correct choice: Instance segmentation
Why: Separates each pill individually, enabling accurate counting even when pills touch