Computer Vision Interview Questions (2026)
Practice 45 Computer Vision interview questions on image tensors, CNNs, augmentation, transfer learning, object detection, segmentation, metrics, OpenCV, and deployment.
45 questions with answers- What Is Computer Vision?
- All Questions on This Page
- Computer Vision Fundamentals
- Computer Vision Practical Interview Questions
- Computer Vision Advanced Scenarios
- Classification vs Detection vs Segmentation
- How to Prepare for a Computer Vision Interview
- How Strong Computer Vision Answers Sound
- Computer Vision Quiz
- Frequently Asked Questions
What Is Computer Vision?
Key Takeaways
- ✓Computer vision uses algorithms and machine learning models to interpret images, video, and visual signals.
- ✓Interviews test image tensors, preprocessing, CNNs, transfer learning, detection, segmentation, tracking, metrics, and deployment constraints.
- ✓Strong answers connect model choice with data quality, labels, edge cases, latency, and evaluation metric.
- ✓Vision projects fail when candidates ignore lighting, camera angle, label noise, class imbalance, and production image drift.
Computer vision is the field of building systems that understand images and video. In interviews, computer vision questions test how you represent images as tensors, preprocess and augment data, train CNN or transformer-based models, evaluate detection or segmentation outputs, and deploy models where latency, camera quality, and real-world drift matter.
Watch: TensorFlow for Computer Vision course
Video: TensorFlow for Computer Vision course (freeCodeCamp.org, YouTube)
Test yourself and earn a certificate
6 quick questions. Score 70%+ to download your Computer Vision certificate.
All Questions on This Page
45 questions- 1. How is an image represented for a model?
- 2. Why do RGB and BGR channel orders matter?
- 3. Why normalize images?
- 4. What does a convolution do?
- 5. What is a CNN?
- 6. Why use pooling?
- 7. What is image augmentation?
- 8. Why use transfer learning?
- 9. What is object detection?
- 10. What is Intersection over Union?
- 11. What is mAP in object detection?
- 12. What is image segmentation?
- 13. What is OCR?
- 14. What does NMS do?
- 15. What does drift mean in vision?
- 16. Design an image classifier for defects.
- 17. How do you choose augmentations?
- 18. How do you fine-tune a pretrained vision model?
- 19. How would you build a helmet detector?
- 20. How would you build a road segmentation model?
- 21. How do you design an OCR pipeline for forms?
- 22. Where does OpenCV fit in a vision pipeline?
- 23. Which metric do you use for object detection?
- 24. How do you handle rare visual classes?
- 25. How do you deploy vision on an edge device?
- 26. How do you split image data correctly?
- 27. How do you reduce false positives?
- 28. How is video different from image inference?
- 29. How do you explain a vision model's mistakes?
- 30. How do you monitor a deployed vision model?
- 31. A new camera reduces model accuracy. What happened?
- 32. Accuracy is high but users complain. What do you check?
- 33. Detection mAP drops after deployment. What do you inspect?
- 34. A detector returns many boxes for one object. What do you tune?
- 35. A model misses small objects. What changes help?
- 36. Segmentation masks are poor near edges. What do you inspect?
- 37. Annotators disagree on labels. What do you do?
- 38. Augmentation makes validation worse. Why?
- 39. A model is accurate but too slow on device. What do you try?
- 40. OCR fails on scanned forms. What helps?
- 41. Vision metrics look too good. What do you suspect?
- 42. The model fails at night. What is missing?
- 43. Predicted labels are swapped. What do you check?
- 44. A face detection system raises privacy concerns. What should be discussed?
- 45. What do you ask before approving a vision model?
Computer Vision Fundamentals
Foundational15 questionsStart here. These are the definitions and first-principle checks that open most rounds.
Q1. How is an image represented for a model?
An image is represented as a tensor with height, width, channels, and often a batch dimension.
Frameworks may use channel-last or channel-first layout, so shape matters.
Image tensor changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Watch a deeper explanation
Video: TensorFlow for Computer Vision course (freeCodeCamp.org, YouTube)
Q2. Why do RGB and BGR channel orders matter?
RGB and BGR place color channels in different order, so wrong ordering changes the image the model sees.
OpenCV often uses BGR while many deep learning models expect RGB.
Color channels changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q3. Why normalize images?
Normalization puts pixel values on a scale that makes optimization easier and matches pretrained model expectations.
Use the same normalization at training and inference.
Normalization changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q4. What does a convolution do?
A convolution applies learned filters over local image regions to detect patterns such as edges, textures, and shapes.
Stacked convolutions build higher-level features.
Convolution changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Vision model path
The head changes with task type.
| Answer part | What to say | Evidence to mention |
|---|---|---|
| Definition | Convolution in one direct sentence. | Official docs or course material |
| Use case | The work where it changes a decision. | Dataset, model, query, dashboard, or pipeline |
| Risk | What breaks when it is misunderstood. | Metric, log, test result, or review note |
Q5. What is a CNN?
A convolutional neural network uses convolution, activation, pooling, and learned layers to process image data.
CNNs remain common for image classification and many embedded use cases.
CNN changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Watch a deeper explanation
Video: PyTorch computer vision (PyTorch tutorial, YouTube)
Q6. Why use pooling?
Pooling reduces spatial size and summarizes local features.
It can reduce compute, but too much pooling loses detail.
Pooling changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q7. What is image augmentation?
Augmentation creates training variations such as crops, flips, rotations, brightness changes, or blur.
Use augmentations that could happen in the real world.
Data augmentation changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q8. Why use transfer learning?
Transfer learning starts from a pretrained vision model and adapts it to a new dataset.
It helps when labeled data is limited.
Transfer learning changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q9. What is object detection?
Object detection predicts bounding boxes and class labels for objects in an image.
Evaluation usually uses IoU and mAP.
Object detection changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q10. What is Intersection over Union?
IoU measures overlap between predicted and ground-truth boxes or masks.
It is intersection area divided by union area.
IoU changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q11. What is mAP in object detection?
mAP summarizes precision-recall performance across classes and IoU thresholds.
It is more informative than raw accuracy for detection.
mAP changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q12. What is image segmentation?
Segmentation assigns labels to pixels or predicts object masks.
Semantic segmentation labels classes. Instance segmentation separates object instances.
Segmentation changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q13. What is OCR?
OCR extracts text from images or scanned documents.
Real OCR systems need detection, recognition, language rules, and post-processing.
OCR changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Watch a deeper explanation
Video: OpenCV Python tutorial (OpenCV tutorial, YouTube)
Q14. What does NMS do?
NMS removes duplicate detection boxes by keeping high-confidence boxes and suppressing overlapping lower-confidence boxes.
Bad NMS settings can remove true objects or keep duplicates.
Non-maximum suppression changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q15. What does drift mean in vision?
Drift means production images differ from training images due to camera, lighting, background, product, or user behavior changes.
Vision systems need monitoring and sample review.
Model drift changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Computer Vision Practical Interview Questions
Intermediate15 questionsThese questions test whether you can apply the topic to real data, real code, and messy constraints.
Q16. Design an image classifier for defects.
Define classes, collect labeled examples, split by source, resize and normalize images, train a baseline, evaluate per class, and inspect mistakes.
Defect datasets often have imbalance and subtle labels.
image classifier changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
# Interview check: split data, run baseline, inspect metric
from sklearn.metrics import accuracy_score
print(accuracy_score(y_true, y_pred))Q17. How do you choose augmentations?
Choose augmentations that match real variation: lighting, crop, blur, rotation, viewpoint, or background changes.
Do not use flips or rotations that change label meaning.
augmentation policy changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q18. How do you fine-tune a pretrained vision model?
Freeze the base, train a task head, unfreeze selected layers with a lower learning rate, and validate on realistic images.
Track class-level errors, not only top-line score.
transfer learning plan changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q19. How would you build a helmet detector?
Label boxes, train a detector, tune confidence and NMS thresholds, evaluate mAP and false positives, then test on site-specific footage.
Camera angle and occlusion matter.
object detection project changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q20. How would you build a road segmentation model?
Collect pixel masks, train a segmentation model, evaluate mIoU or Dice, and inspect boundary errors.
Thin objects and class imbalance are common pain points.
segmentation project changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q21. How do you design an OCR pipeline for forms?
Detect text regions, correct orientation, recognize text, validate fields, and route low-confidence cases to review.
OCR needs business validation after recognition.
OCR pipeline changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q22. Where does OpenCV fit in a vision pipeline?
Use OpenCV for image loading, resizing, color conversion, filtering, thresholding, contours, and geometry before or after model inference.
Keep preprocessing identical between training and inference.
OpenCV preprocessing changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q23. Which metric do you use for object detection?
Use IoU-based metrics such as mAP, plus precision and recall at operating thresholds.
Choose threshold based on false-positive and false-negative cost.
metric choice changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q24. How do you handle rare visual classes?
Collect more examples, use sampling, class-aware loss, augmentation, threshold tuning, and per-class reporting.
A model can ignore rare classes and still look good.
class imbalance changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q25. How do you deploy vision on an edge device?
Pick a smaller model, resize inputs, quantize if fit, batch carefully, and measure latency on the target hardware.
Laptop latency does not predict edge latency.
latency budget changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Watch a deeper explanation
Video: Convolutional neural networks explained (3Blue1Brown, YouTube)
Q26. How do you split image data correctly?
Split by source, patient, camera, user, site, or time when leakage risk exists.
Near-duplicate images across splits inflate metrics.
data split changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q27. How do you reduce false positives?
Inspect false positives, tune thresholds, add hard negatives, improve labels, and adjust post-processing.
The right threshold depends on the product cost.
false positive review changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q28. How is video different from image inference?
Use frame sampling, tracking, temporal smoothing, latency controls, and drift checks.
Frame-by-frame predictions can flicker.
video inference changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q29. How do you explain a vision model's mistakes?
Review examples by class, camera, lighting, confidence, and confusion pairs; use saliency only as supporting evidence.
Human image review is still essential.
explain model errors changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q30. How do you monitor a deployed vision model?
Track confidence, class mix, input quality, latency, review samples, drift signals, and downstream business outcomes.
Images can drift without schema changes.
monitor production changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Computer Vision Advanced Scenarios
Advanced15 questionsAdvanced rounds test trade-offs, failure modes, and whether the decision can hold up under production pressure.
Q31. A new camera reduces model accuracy. What happened?
Input distribution changed: resolution, color, lens, angle, compression, or lighting may differ.
Add camera-specific validation and retraining data.
camera changed changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q32. Accuracy is high but users complain. What do you check?
Check class imbalance, thresholds, false positives, false negatives, and real user cost.
Top-line accuracy can hide painful errors.
high accuracy low value changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q33. Detection mAP drops after deployment. What do you inspect?
Check image quality, object size, occlusion, label policy, threshold, NMS, and new object appearances.
Detection is sensitive to real-world visuals.
mAP drops changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q34. A detector returns many boxes for one object. What do you tune?
Tune NMS threshold, confidence threshold, anchor settings if applicable, and training labels.
Post-processing matters.
duplicate boxes changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q35. A model misses small objects. What changes help?
Use higher input resolution, feature pyramid models, better labels, tiling, or task-specific architecture.
Small objects disappear after aggressive resizing.
small objects missed changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q36. Segmentation masks are poor near edges. What do you inspect?
Check mask label quality, resolution, loss function, class imbalance, and boundary examples.
Pixel labels are expensive and often noisy.
bad masks changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q37. Annotators disagree on labels. What do you do?
Define labeling rules, measure agreement, resolve edge cases, and keep an audit set.
Unclear labels cap model quality.
label disagreement changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q38. Augmentation makes validation worse. Why?
The augmentation may create unrealistic images or change the label.
Match augmentation to real variation.
augmentation hurts changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q39. A model is accurate but too slow on device. What do you try?
Use smaller architecture, lower resolution, quantization, pruning, batching changes, or hardware acceleration.
Measure on the target device.
edge latency fail changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q40. OCR fails on scanned forms. What helps?
Improve deskewing, denoising, contrast, text detection, field validation, and manual review for low confidence.
OCR is rarely just recognition.
OCR messy scans changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q41. Vision metrics look too good. What do you suspect?
Near-duplicate frames or same source images may exist in both train and validation sets.
Split by source or time.
train test leakage changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q42. The model fails at night. What is missing?
Training data likely lacks night lighting examples or suitable augmentation.
Collect representative samples.
night images fail changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q43. Predicted labels are swapped. What do you check?
Check class index mapping, dataset loader, saved metadata, and post-processing label names.
Label maps should ship with the model.
class map bug changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q44. A face detection system raises privacy concerns. What should be discussed?
Discuss consent, retention, anonymization, access, purpose limits, and legal requirements.
Vision systems often process sensitive data.
privacy concern changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Q45. What do you ask before approving a vision model?
Ask about label policy, split method, metrics, thresholds, edge cases, deployment latency, monitoring, and human review path.
These decide whether the model survives production.
senior review changes Computer Vision decisions by affecting the chosen method, the failure mode, or the validation step. Reliable evidence comes from a metric, test result, query output, log, or review note.
Classification vs Detection vs Segmentation
Vision interviews often start by matching the task to the output format. A classifier predicts an image label. A detector predicts boxes. A segmentation model predicts pixels or masks.
| Task | Output | Common metric | Example |
|---|---|---|---|
| Classification | Class label or probabilities | Accuracy, F1, AUC | Is this image defective? |
| Object detection | Bounding boxes plus classes | mAP, IoU | Find helmets in a worksite image |
| Semantic segmentation | Class per pixel | mIoU, Dice | Mark road, car, and sidewalk pixels |
| Instance segmentation | Separate object masks | Mask mAP | Separate each person in a frame |
Computer vision answer signals
Good answers model output maps to labels and metrics.
- Data quality: labels, lighting, camera, drift
- Task fit: classification, detection, segmentation
- Evaluation: IoU, mAP, Dice, confusion matrix
- Deployment: latency, edge, monitoring
How to Prepare for a Computer Vision Interview
Prepare by tracing one image through the whole system: capture, preprocessing, model, post-processing, metric, and production monitoring.
- Review image shapes, channels, normalization, resizing, augmentation, CNN layers, and transfer learning.
- Know classification, detection, segmentation, OCR, tracking, and feature extraction use cases.
- IoU, mAP, Dice, confusion matrix, precision, recall, and threshold tuning is the explanation path.
- One story about bad labels, camera change, lighting shift, class imbalance, or deployment latency is useful.
Computer vision interview prep flow
Most rounds reward clear reasoning more than memorized phrasing.
How Strong Computer Vision Answers Sound
A strong computer vision answer starts with the visual task and label format. Say whether the problem needs classification, detection, segmentation, OCR, or tracking; describe image size, channels, preprocessing, augmentation, model family, metric, and failure cases. production-ready answers mention data collection bias, annotation quality, calibration, edge latency, false-positive cost, and monitoring camera drift after deployment.
Start with the label
The label format decides the model family and metric. Boxes, masks, and class labels are different products.
Name visual failure modes
Lighting, blur, occlusion, viewpoint, background, resolution, and camera change can break a model.
Use the right metric
Accuracy is not enough for detection or segmentation. Use IoU, mAP, Dice, precision, and recall where they fit.
Plan for deployment
Vision models often run with latency, memory, power, and privacy constraints.
Test Yourself: Computer Vision Quiz
Ready to test your Computer Vision knowledge?
6 questions, about 4 minutes. Score 70% or higher to earn a shareable certificate.
Frequently Asked Questions
What do Computer Vision interviews usually ask?
Do I need OpenCV for Computer Vision interviews?
What Computer Vision project should I discuss?
What separates experienced Computer Vision candidates?
Practice vision answers before ML interviews
Use this Computer Vision bank for model and evaluation reasoning, then practice Python tasks with Hyring's AI Coding Interviewer.
See Hyring AI Coding InterviewerSources
Written by Adithyan RK