Computer vision interview questions with answer outlines

Fifteen common computer vision interview questions

1. How does a convolution work, and why share weights?
2. What do pooling layers and strides do?
3. What is a receptive field?
4. Why was ResNet a breakthrough?
5. Compare two-stage and one-stage object detectors.
6. What is intersection over union (IoU)?
7. What is non-maximum suppression?
8. How is mean average precision (mAP) calculated?
9. Semantic versus instance segmentation?
10. How does U-Net work?
11. Which data augmentations would you use, and which would you avoid?
12. How do vision transformers process images?
13. How do you handle class imbalance in detection?
14. How would you deploy a model on an edge device?
15. Design a system to detect defects on a production line.

CNN fundamentals

• Convolution: a small filter slides over the image computing weighted sums; the same weights are reused everywhere, so a feature such as an edge is detected regardless of position, with far fewer parameters than a dense layer.
• Pooling and stride: reduce spatial size, adding some translation tolerance and cutting computation.
• Receptive field: the input region influencing one output unit; it grows with depth, letting deep layers capture objects and context.
• ResNet: residual connections let gradients flow through very deep networks, making 50- to 150-layer models trainable and accurate.

Object detection

Segmentation

Semantic segmentation labels every pixel with a class but doesn't separate individual objects; instance segmentation separates each object; panoptic segmentation combines both. U-Net uses an encoder that downsamples to capture context and a decoder that upsamples, with skip connections passing fine detail across, which is why it works well with limited data such as medical images. Mask R-CNN extends a two-stage detector with a mask branch per detected object.

Data, augmentation and imbalance

• Augmentation: flips, crops, scaling, colour jitter, blur and cut-and-mix methods — but only those that preserve the label (don't flip text or left/right-specific medical images).
• Labelling quality often matters more than architecture; audit a sample and define clear guidelines.
• Imbalance: focal loss down-weights easy examples; oversample rare classes; evaluate per class.

Vision transformers and deployment

Vision transformers split an image into patches, embed each as a token and apply self-attention, capturing global relationships; they excel with large pre-training datasets, while CNNs remain efficient on smaller data and edge hardware. For edge deployment, use smaller architectures, quantise to 8-bit integers, prune, and use hardware-specific runtimes, measuring accuracy after each step. See deep learning interview questions.

A design question: defect detection

Clarify defect types, frequency, line speed and the cost of misses versus false alarms. Collect images under controlled lighting, label defects with boxes or masks, and expect severe imbalance since defects are rare — consider anomaly detection trained mostly on good parts. Choose a fast detector that meets the line's frame rate on edge hardware, set the threshold from the cost trade-off, route uncertain cases to human review, and monitor drift when lighting, cameras or products change.

Practising

Interviewers follow up on the most specific thing you say, so rehearse the second question as well as the first. CoPilot Interview, a free AI interview assistant, runs computer vision mock interviews with follow-ups and produces a report of likely next-round questions. Prepare to explain one vision project end to end, including how you labelled data and which augmentations you rejected.

Disclosure: CoPilot Interview is our product, so weigh what we say about it accordingly. Everything else on this page applies whichever tool you use, or none.

Frequently asked questions

What are common computer vision interview questions?

Convolutions and CNN basics, ResNet, two-stage versus one-stage detectors, IoU, non-maximum suppression, mAP, semantic versus instance segmentation, U-Net, augmentation, vision transformers and edge deployment.

What is non-maximum suppression?

A post-processing step that keeps the highest-confidence bounding box and removes other boxes that overlap it beyond an IoU threshold, so each object is detected once.

YOLO or Faster R-CNN?

One-stage detectors like YOLO are faster and suit real-time use. Two-stage detectors like Faster R-CNN are often more accurate, especially for small objects, at higher computational cost.

How do vision transformers differ from CNNs?

They split images into patches treated as tokens and use self-attention to model global relationships. They need more pre-training data, while CNNs are more efficient on small datasets and edge devices.