Semantic segmentation vs instance segmentation in Computer Vision - Practice Questions

Challenge - 5 Problems

🧠 Conceptual

intermediate

Difference between semantic and instance segmentation

Which statement correctly describes the difference between semantic segmentation and instance segmentation?

A Semantic segmentation detects objects with bounding boxes, while instance segmentation labels pixels with classes.

B Instance segmentation only works on images with a single object, semantic segmentation works on multiple objects.

C Instance segmentation labels each pixel with a class, while semantic segmentation distinguishes different objects of the same class.

D Semantic segmentation labels each pixel with a class, while instance segmentation also distinguishes different objects of the same class.

💡 Hint

Think about whether the method separates individual objects or just labels the type of object.

Semantic segmentation assigns a class label to every pixel but does not separate different objects of the same class. Instance segmentation does both: it labels pixels and separates individual objects.

❓ Predict Output

intermediate

Output shape of semantic vs instance segmentation masks

Given an input image of size 256x256 pixels, what is the shape of the output mask for semantic segmentation and instance segmentation respectively?

A Semantic: (256, 256), Instance: (256, 256)

B Semantic: (256, 256, num_classes), Instance: (num_instances, 256, 256)

C Semantic: (num_instances, 256, 256), Instance: (256, 256, num_classes)

D Semantic: (256, 256, 3), Instance: (256, 256, 3)

💡 Hint

Semantic segmentation outputs a class label per pixel; instance segmentation outputs masks per object.

Semantic segmentation outputs a mask with one channel per class, so shape is (height, width, num_classes). Instance segmentation outputs a mask per detected object, so shape is (num_instances, height, width).

❓ Model Choice

advanced

Choosing a model for instance segmentation

You want to detect and separate multiple objects of the same class in images. Which model architecture is best suited for this task?

A Mask R-CNN

B ResNet for classification

C Fully Convolutional Network (FCN)

D U-Net

💡 Hint

Look for a model that can detect objects and segment them individually.

Mask R-CNN performs instance segmentation by detecting objects and generating masks for each instance. U-Net and FCN are mainly for semantic segmentation. ResNet is for classification.

❓ Metrics

advanced

Evaluating semantic vs instance segmentation

Which metric is commonly used to evaluate both semantic and instance segmentation performance?

A Mean Intersection over Union (mIoU)

B Mean Squared Error (MSE)

C BLEU score

D Accuracy of bounding box coordinates

💡 Hint

Think about a metric that compares predicted and true pixel areas.

Mean Intersection over Union (mIoU) measures overlap between predicted and true masks and is widely used for both semantic and instance segmentation. Bounding box accuracy is for detection, BLEU is for language, MSE is for regression.

🔧 Debug

expert

Debugging instance segmentation mask output

You trained an instance segmentation model but the output masks overlap heavily, merging different objects into one mask. What is the most likely cause?

A The input images are too small, causing poor resolution.

B The learning rate is too low, causing slow convergence.

C The model's non-maximum suppression (NMS) threshold is too high, causing overlapping masks to merge.

D The model uses semantic segmentation instead of instance segmentation.

💡 Hint

Consider how overlapping predictions are filtered in instance segmentation.

Non-maximum suppression (NMS) removes overlapping detections. If its threshold is too high, masks that should be separate merge. Learning rate and image size affect training quality but not mask merging. Using semantic segmentation would not produce instance masks.