Computer Vision interview questions

26 questions from the Computer Vision course, each with the correct answer and an explanation. Try to answer before you open the answer.

Images and Vision Tasks

1. Which task outputs bounding boxes with labels?

  1. Semantic segmentation
  2. Image classification
  3. Object detection
  4. OCR only
Show answer

Answer: Object detection

Detection localises objects with boxes.

Lesson: What Computer Vision Does

2. What layout does PyTorch expect for an image tensor?

  1. (channels, height, width)
  2. (height, width, channels)
  3. (width, height)
  4. (pixels, labels)
Show answer

Answer: (channels, height, width)

Convert with permute when moving from NumPy.

Lesson: Images as Arrays of Pixels

3. Why is HSV useful for colour selection?

  1. It removes all noise
  2. It uses fewer pixels
  3. It is the format neural networks require
  4. It separates colour (hue) from brightness, so masks are more robust to lighting
Show answer

Answer: It separates colour (hue) from brightness, so masks are more robust to lighting

Hue stays similar when brightness changes.

Lesson: Colour Spaces

Classical Image Processing

4. What is a trade-off of Gaussian blurring?

  1. It sharpens every edge
  2. It reduces noise but also softens edges
  3. It changes image size
  4. It converts the image to colour
Show answer

Answer: It reduces noise but also softens edges

Smoothing removes fine detail too.

Lesson: Filtering and Denoising

5. What happens when you raise Canny's thresholds?

  1. More edges appear
  2. Fewer, stronger edges are kept
  3. The image becomes coloured
  4. Edges are blurred
Show answer

Answer: Fewer, stronger edges are kept

Higher thresholds keep only strong gradients.

Lesson: Edge Detection

6. Why did one coin region come out much larger than the others?

  1. Labels were assigned randomly
  2. Otsu always fails
  3. The image is in colour
  4. Several touching coins merged into one connected component
Show answer

Answer: Several touching coins merged into one connected component

Touching objects need separation methods.

Lesson: Thresholding and Connected Components

7. What does RANSAC do in feature matching?

  1. Fits a geometric transform while rejecting outlier matches
  2. Detects corners
  3. Blurs the image
  4. Classifies objects
Show answer

Answer: Fits a geometric transform while rejecting outlier matches

Robust fitting ignores wrong matches.

Lesson: Keypoints and Feature Matching

Geometry, Augmentation and Data

8. What must happen to bounding boxes when an image is rotated for training?

  1. They should be deleted
  2. Nothing
  3. They must be transformed with the same rotation
  4. They should be doubled in size
Show answer

Answer: They must be transformed with the same rotation

Labels must follow the pixels.

Lesson: Geometric Transforms

9. Why is test-time preprocessing kept deterministic?

  1. Tests have no images
  2. Augmentation is illegal at test time
  3. So evaluation scores are stable and comparable
  4. It makes training faster
Show answer

Answer: So evaluation scores are stable and comparable

Random changes belong in training.

Lesson: Data Augmentation

10. Why should frames from the same video not be split across train and test?

  1. Near-duplicate frames inflate test scores
  2. Videos cannot be labelled
  3. Frames are always blank
  4. It reduces training speed
Show answer

Answer: Near-duplicate frames inflate test scores

Keep related images on one side of the split.

Lesson: Datasets and Labelling

Image Classification With Deep Learning

11. What made very deep CNNs like ResNet trainable?

  1. Training without labels
  2. Removing all convolutions
  3. Using grayscale only
  4. Residual (skip) connections
Show answer

Answer: Residual (skip) connections

Skip connections keep gradients flowing.

Lesson: CNN Architectures in Brief

12. Why must you use the same preprocessing as the pretrained weights?

  1. It makes images larger
  2. The model learned from inputs scaled and normalised that way
  3. Any preprocessing works the same
  4. It is only needed for training
Show answer

Answer: The model learned from inputs scaled and normalised that way

Mismatched inputs degrade predictions.

Lesson: Classifying With a Pretrained Model

13. What does a linear probe train?

  1. Every layer of the network from scratch
  2. Only a simple classifier on top of frozen pretrained features
  3. The image sensor
  4. A new tokenizer
Show answer

Answer: Only a simple classifier on top of frozen pretrained features

Cheap, fast and often strong.

Lesson: Transfer Learning: Linear Probes and Fine-Tuning

Object Detection

14. A prediction shifted by 20 pixels on a 100-pixel box has IoU of about...

  1. 2.0
  2. 1.0
  3. 0.0
  4. 0.67
Show answer

Answer: 0.67

Overlap shrinks quickly with shifts.

Lesson: Bounding Boxes and IoU

15. What does NMS remove?

  1. Lower-scoring boxes that heavily overlap a higher-scoring box
  2. All boxes below 0.99
  3. Every other box
  4. Ground-truth labels
Show answer

Answer: Lower-scoring boxes that heavily overlap a higher-scoring box

It removes duplicates of the same object.

Lesson: Non-Maximum Suppression

16. Why apply a score threshold to detector outputs?

  1. Thresholds make boxes larger
  2. Scores are always 1
  3. Detectors return many low-confidence boxes that are usually wrong
  4. It is needed to load the model
Show answer

Answer: Detectors return many low-confidence boxes that are usually wrong

Raw outputs include many weak guesses.

Lesson: Running a Pretrained Detector

17. Why is a second box on an already-matched object counted as a false positive?

  1. It has a low IoU
  2. Boxes cannot overlap
  3. Each ground-truth object can be matched only once
  4. All second boxes are correct
Show answer

Answer: Each ground-truth object can be matched only once

Duplicates are penalised.

Lesson: Evaluating Detectors: Precision, Recall and AP

Segmentation

18. Which segmentation type separates two touching people into different masks?

  1. Instance segmentation
  2. Semantic segmentation only
  3. Image classification
  4. Edge detection
Show answer

Answer: Instance segmentation

Instances are individual objects.

Lesson: Semantic, Instance and Panoptic Segmentation

19. How is the final class map usually obtained from a segmentation model?

  1. Using the image's file name
  2. Averaging all pixels
  3. Choosing a random class
  4. Taking the argmax over class scores for each pixel
Show answer

Answer: Taking the argmax over class scores for each pixel

One class per pixel.

Lesson: Running a Pretrained Segmentation Model

20. How do Dice and IoU compare for the same masks?

  1. Dice is always lower
  2. Dice is always greater than or equal to IoU
  3. They are identical
  4. They measure different images
Show answer

Answer: Dice is always greater than or equal to IoU

Do not compare one with the other.

Lesson: Segmentation Metrics: IoU and Dice

Robustness and Deployment

21. Why track probabilities, not only the top label, in robustness tests?

  1. Probabilities never change
  2. Confidence can collapse even when the top label stays the same
  3. Top labels are always wrong
  4. It makes models faster
Show answer

Answer: Confidence can collapse even when the top label stays the same

Falling confidence is an early warning.

Lesson: Robustness to Real-World Conditions

22. What does cv2.imread return by default?

  1. Grayscale only
  2. Pixels in RGB order
  3. A PyTorch tensor
  4. Pixels in BGR channel order
Show answer

Answer: Pixels in BGR channel order

Convert to RGB if your model expects it.

Lesson: Preprocessing Consistency

23. Why deploy a vision model on an edge device?

  1. Lower latency, offline operation and keeping images on site
  2. To make models larger
  3. Because cloud GPUs cannot run CNNs
  4. To avoid any testing
Show answer

Answer: Lower latency, offline operation and keeping images on site

Edge trades capacity for latency and privacy.

Lesson: Deploying Vision Models

Foundation Models, Ethics and a Checklist

24. What does zero-shot classification with CLIP-style models need?

  1. A segmentation mask
  2. Thousands of labelled images per class
  3. Text descriptions of the classes, not labelled training images
  4. A new GPU architecture
Show answer

Answer: Text descriptions of the classes, not labelled training images

Images are compared with text prompts.

Lesson: Vision Foundation Models

25. Which practice reduces privacy risk in a people-counting system?

  1. Sharing images widely for debugging
  2. Storing all raw video forever
  3. Counting on the device and discarding or blurring faces
  4. Adding face recognition by default
Show answer

Answer: Counting on the device and discarding or blurring faces

Collect and keep only what you need.

Lesson: Privacy, Fairness and Responsible Use

26. Which item belongs on a computer vision launch checklist?

  1. Identical preprocessing in training and serving, verified with golden images
  2. Feeding BGR images to an RGB model
  3. Splitting video frames randomly across train and test
  4. Skipping latency tests
Show answer

Answer: Identical preprocessing in training and serving, verified with golden images

Consistency and realistic testing.

Lesson: A Computer Vision Project Checklist

← Back to the course