Q1
Walk me through how you would design a real-time object detection pipeline using a pre-trained model like YOLOv8 or Faster R-CNN. What considerations would you make for inference latency, memory usage, and accuracy trade-offs in a production environment?
Why they ask this:* They want to assess your understanding of end-to-end computer vision system design, optimization techniques, and practical deployment constraints beyond just model selection.
Q2
Explain the differences between transfer learning and fine-tuning for image classification tasks. When would you choose one approach over the other, and what are the risks of overfitting in each scenario?
Why they ask this:* This tests your grasp of fundamental CV training strategies, your ability to optimize model performance with limited data, and your understanding of when to adapt existing models versus training from scratch.
Q3
You have a dataset with significant class imbalance (90% background, 10% objects of interest) for a segmentation task. What techniques would you implement to address this, and how would you choose appropriate evaluation metrics?
Why they ask this:* They're evaluating your practical experience with real-world data challenges, your knowledge of data augmentation and sampling strategies, and your ability to select metrics beyond accuracy.
Q4
Describe your experience with deep learning frameworks (PyTorch, TensorFlow) for computer vision tasks. Can you walk through your typical workflow for debugging a model that achieves good training accuracy but poor validation performance?