What you need to know
Practise ten Computer Vision Engineer interview questions on task choice, geometry, dataset leakage, annotation quality, IoU, confidence, tracking and real-time deployment.
Practise ten Computer Vision Engineer interview questions on task choice, geometry, dataset leakage, annotation quality, IoU, confidence, tracking and real-time deployment.
These are original practice questions, not questions reported from a particular employer. Every scenario below is hypothetical. Use the answers to understand the reasoning, then explain the decision in your own words and connect it to work you actually did.
How to practise
First answer aloud without reading. State the goal, assumptions, proposed approach, checks and trade-offs. Then compare your reasoning with the answer, tackle the follow-up and write down one gap to practise. Aim for a clear two-minute initial explanation; expand when asked. This timing is a practice suggestion, not an interview rule.
Role-based questions and detailed answers
1. How do classification, detection and segmentation differ?
Answer: Classification assigns a label to an image or region, detection locates objects with bounding boxes, and segmentation assigns labels at pixel level. Choose the output that supports the actual decision and available annotation budget. More detailed output is not automatically better; it increases labeling and evaluation requirements and may add complexity without improving the user's outcome.
Hypothetical scenario: A warehouse needs to count individual boxes rather than merely determine whether any box appears in a frame.
Follow-up: Would image classification provide enough information for that decision?
Common mistake: Choosing the most complex vision task without matching the required output.
2. Why must coordinate transforms be tracked during preprocessing?
Answer: Resizing, cropping and padding change how model coordinates map to the original image. Record the exact transformation and invert it consistently when displaying boxes or using spatial measurements. Test known points and boundary cases, including non-square inputs. A correct detector can appear wrong if postprocessing uses the wrong scale, offset or coordinate convention.
Hypothetical scenario: A rectangular image is padded to a square, but output boxes are rescaled without subtracting padding.
Follow-up: Which simple fixture would reveal the offset error?
Common mistake: Assuming width and height scale equally after every preprocessing operation.
3. Why split visual data by capture session?
Answer: Adjacent frames and repeated views often share background, lighting and object identity. Randomly splitting them can place near-duplicates in training and evaluation, overstating performance on genuinely new conditions. Group by session, object or site according to the deployment claim. Keep difficult conditions represented without leaking the same scene across partitions under different filenames.
Hypothetical scenario: A video contributes consecutive frames to both training and test sets.
Follow-up: Which grouping would better test performance at a new location?
Common mistake: Treating different image files as automatically independent samples.
4. How do you handle ambiguous annotations?
Answer: Write explicit labeling rules for occlusion, truncation, uncertain classes and overlapping objects. Review disagreements and update guidance before scaling annotation. Preserve uncertainty where the task permits it and audit examples by source and condition. A model cannot be evaluated meaningfully against inconsistent targets without acknowledging that part of the apparent error may come from annotation policy.
Hypothetical scenario: Annotators disagree on whether a mostly hidden object should receive a full box or only a visible-region box.
Follow-up: How would you make that convention consistent across the dataset?
Common mistake: Blaming the model for disagreement caused by undefined labeling rules.
5. How do you calculate intersection over union?
Answer: IoU divides the overlapping area of two regions by their union area. If boxes have areas 100 and 80 and overlap by 60, the union is 100 plus 80 minus 60, or 120, so IoU is 0.5. Interpret the result within the task's matching rules; IoU alone does not establish correct class, counting or overall system quality.
Hypothetical scenario: A predicted box covers the correct object but also includes a substantial portion of the background.
Follow-up: Why can a correct class label still produce a poor localization score?
Common mistake: Dividing overlap by only the predicted box area and calling it IoU.
6. Does a high detection confidence guarantee correctness?
Answer: No. The score reflects the model and its training, not a universal probability guarantee under every condition. Evaluate score behaviour on representative data, including unfamiliar backgrounds and classes. Choose thresholds according to missed-detection and false-alarm costs, and inspect calibration where probabilities matter. A confident prediction under distribution shift can still be completely wrong.
Hypothetical scenario: A detector confidently labels a printed photograph as a physical object in a live scene.
Follow-up: Which evaluation slice would expose this failure?
Common mistake: Treating a large confidence score as independent verification of reality.
7. How are duplicate suppression and tracking different?
Answer: Duplicate suppression attempts to reduce overlapping detections of the same object within a frame. Tracking associates objects across frames to maintain identity over time. Neither alone guarantees accurate unique counts, especially with occlusion or re-entry. Define the counting rule, evaluate identity switches and inspect where suppression incorrectly removes nearby distinct objects.
Hypothetical scenario: A camera counts the same passing object once in every frame despite clean per-frame detections.
Follow-up: What additional state is needed to count a passage once?
Common mistake: Assuming non-maximum suppression solves cross-frame identity and counting.
8. What can synthetic image stress tests tell you?
Answer: Controlled changes such as blur or brightness can expose sensitivities and help compare candidate behaviour. They do not establish performance under all real camera, weather or demographic conditions. Keep transformation settings documented and complement them with representative field data collected appropriately. Report them as stress tests rather than claiming they prove robustness in deployment.
Hypothetical scenario: A detector tolerates artificial brightness changes but fails with glare from a real reflective surface.
Follow-up: Which additional data would test the deployment condition directly?
Common mistake: Equating success on synthetic perturbations with proven real-world robustness.
9. Why can a fast model still produce stale video results?
Answer: End-to-end latency includes capture, buffering, decoding, preprocessing, inference and downstream display or actions. If input arrives faster than processing, an unbounded queue can make results increasingly stale despite stable per-frame inference speed. Choose a buffering or frame-dropping policy that matches the task, and measure frame age as well as throughput under realistic load.
Hypothetical scenario: A detector processes ten frames per second while its camera supplies thirty, creating a growing backlog.
Follow-up: When is dropping old frames safer than processing every frame?
Common mistake: Reporting model FPS while ignoring queue growth and observation age.
10. What evidence should accompany a vision deployment?
Answer: Document supported conditions, representative evaluation data, important error slices and the consequences of mistakes. Include camera configuration, preprocessing, latency, monitoring and a fallback or review process where needed. Address image access, retention and privacy for the actual use. A convincing demo is evidence that the pipeline runs, not proof it is suitable for every site or person.
Hypothetical scenario: A prototype works on a developer's webcam but is proposed for an unattended production environment.
Follow-up: What must be tested before expanding beyond the demonstration conditions?
Common mistake: Generalizing from a short demo without defining supported operating conditions.
Turn practice into evidence
Build an object-detection demonstration using images you own or have permission to use. Document annotation rules, split by capture session and test coordinate transformations after resizing. Report errors by lighting and object size, and measure frame freshness as well as inference speed. Treat stress tests as controlled probes, not measured field performance.
Score each response informally: 0 if you cannot explain it, 1 for a definition, 2 for an approach with checks, and 3 for justified trade-offs plus a failure-recovery plan. This is a self-study rubric, not a validated hiring assessment. Revisit weak answers and never invent project outcomes or experience.
Career roadmap and related jobs
Build the underlying skills with the Computer Vision Engineer career guide. For a shorter starting point, read AI interview answer frameworks.
Explore relevant live jobs. Listings are filtered by Computer Vision and can include adjacent roles. Read the employer's requirements and original posting before applying. Use one listing to select the questions most relevant to its responsibilities.
Methodology
Editorial synthesis of cited sourcesAI-assisted original educational practice material. These are not actual employer interview questions, reported interview experiences or hiring guarantees. Answers, hypothetical scenarios, follow-ups and self-assessment advice are original editorial synthesis. Official Google references support background concepts, not endorsements of these answers. Related jobs use a specialization filter and may include adjacent roles; availability changes. References checked 17 October 2026.
Read our editorial and research policy →Sources
- Object detection task guide
Google AI Edge · Checked 17 October 2026
- Dividing datasets
Google for Developers · Checked 17 October 2026
- Classification: accuracy, recall, precision and related metrics
Google for Developers · Checked 17 October 2026