Computer Vision Engineer Interview Questions UK: 8 Real Questions and What Interviewers Want

Most Computer Vision Engineer interviews in the UK follow a pattern, and knowing it beats guessing. Below are eight questions from our bank of 12, each with a note on what the interviewer is really after.

Eight Computer Vision Engineer interview questions, and what the interviewer is thinking

1. When would you choose a convolutional neural network over a vision transformer for an image task, and vice versa?

Why they ask: A standard architecture-choice question for computer vision roles.

2. How would you optimise a vision model to run in real time on an edge device such as an NVIDIA Jetson?

Why they ask: Deployment optimisation is a daily concern for CV engineers.

3. What is camera calibration, and why does it matter when you need to measure things in the real world from images?

Why they ask: Geometry fundamentals separate CV engineers from general ML practitioners.

4. What is the difference between semantic, instance and panoptic segmentation, and how would you choose loss functions for segmentation tasks?

Why they ask: Tests breadth across core vision tasks and practical training knowledge.

5. You only have a small labelled image dataset. How would you build a reliable model?

Why they ask: Small data is common in real projects, and they want practical strategies.

6. Tell me about a time you improved the speed or efficiency of a computer vision system.

Why they ask: They want evidence of practical engineering impact, not just model accuracy.

7. A retail client asks you to add live facial recognition to its in-store cameras to identify known shoplifters. How would you respond?

Why they ask: UK privacy law and ethics are an essential part of deploying vision systems.

8. What does a typical day as a Computer Vision Engineer involve, and why does the field interest you?

Why they ask: They want to see you understand the balance of engineering and research in the role.

A sample answer to question 1

When would you choose a convolutional neural network over a vision transformer for an image task, and vice versa?

CNNs build in useful inductive biases, locality and translation equivariance, so they learn well from smaller datasets, train quickly and run efficiently, which is why architectures like ResNet, EfficientNet or ConvNeXt remain strong for many production tasks, particularly on constrained hardware. Vision transformers split an image into patches and use self-attention to model global relationships, so they tend to shine when there is a lot of pre-training data or a strong pre-trained backbone, and they scale well to very large models. For a new project I would first try a pre-trained backbone, and which family depends on constraints: for a small labelled dataset on an edge device, I would likely fine-tune a compact CNN or an efficient hybrid; for a larger dataset with complex scenes, where I have GPU budget, a ViT or Swin-style model may deliver better accuracy. I always benchmark candidates on accuracy, latency and memory against the real deployment target rather than relying on leaderboard numbers.

Common mistake to avoid: Declaring transformers always superior; ignoring latency and memory

Want the rest?

Full model answers, key points and common mistakes for all 12 questions are on the Computer Vision Engineer interview questions page. The first five are free, no sign-up needed.