CS5720 - Week 10
Slide 187 of 200

Semantic vs Instance Segmentation

SEMANTIC

Semantic Segmentation

Class-based grouping: All pixels of the same class get the same label, regardless of which specific object instance they belong to.
P
P
B
C
C
B
P
P
P
B
C
B
B
P
P
B
B
B
B
B
B
B
B
B
  • 🎨 All people pixels = green (same color)
  • 🔢 Outputs class probabilities per pixel
  • ⚡ Faster training and inference
  • 🏗️ Simpler network architecture
INSTANCE

Instance Segmentation

Object-based separation: Each individual object instance gets a unique label, distinguishing between different instances of the same class.
P1
P1
B
C1
C1
B
P2
P2
P2
B
C2
B
B
P3
P3
B
B
B
B
B
B
B
B
B
  • 🌈 Each person = different color (P1, P2, P3)
  • 🔍 Can count individual objects
  • 🎯 Precise object boundary detection
  • ⚙️ More complex architecture required

When to Use Which Approach?

SEMANTIC
🚗
Autonomous Driving
Road understanding: drivable area, sidewalks, buildings. Don't need to count individual cars, just know where they are.
INSTANCE
🏭
Factory Quality Control
Count defective products on conveyor belt. Need to identify and track each individual item separately.
SEMANTIC
🌱
Agricultural Monitoring
Crop vs weed classification from satellite images. Focus on area coverage, not individual plants.
INSTANCE
🔬
Cell Biology Research
Count and track individual cells in microscopy images. Each cell needs unique identification.
SEMANTIC
🏥
Medical Imaging
Organ segmentation for surgical planning. Identify tissue types and organ boundaries.
INSTANCE
🤖
Robotics
Object manipulation in cluttered environments. Need to distinguish between individual objects to pick them up.
← Previous Next →
Prepared by Dr. Gorkem Kar