Q1
Walk me through your approach to designing a real-time object detection pipeline for edge devices with limited computational resources. What trade-offs would you consider between model accuracy, latency, and memory footprint?
Why they ask this:* They want to assess your understanding of model optimization, quantization techniques, and practical constraints in production computer vision systems. This demonstrates systems-level thinking beyond just coding.
Q2
Describe your experience with data annotation and labeling strategies for training computer vision models. How would you handle class imbalance, ambiguous edge cases, or noisy labels in a dataset?
Why they ask this:* Data quality directly impacts model performance. They're evaluating your grasp of the full ML pipeline, not just modeling, and your ability to identify and mitigate data-related issues early.
Q3
Explain the architectural differences between CNNs, Vision Transformers, and hybrid models. In what scenarios would you choose each for an image classification task, and what are the computational implications?
Why they ask this:* This tests depth of knowledge in modern architectures and your ability to make informed architectural choices based on problem constraints, not just trending frameworks.
Q4
How do you approach debugging a computer vision model that performs well on validation data but fails on real-world production images? Walk me through your diagnostic process and tools.