Interview Questions
Explaining vision clearly.
Vision interviews test whether you understand trade-offs, and real deployment. Here are the most common questions, with strong answers. Practise them out loud.
- They test trade-offs
- Common questions
- Practise out loud
Design a defect detection system
Start with the requirements. How many items per second? What does a miss cost? Control the lighting first. Then choose a model, usually a small detector or a segmenter with transfer learning. Collect data from the real line, including rare defects, using synthetic data if needed. Measure recall on the defect class. Deploy on the edge, run it in shadow mode, and monitor for drift.
One-stage or two-stage detectors?
Two stage detectors, like Faster RCNN, first propose regions, then classify them. They're often more accurate on small objects, but slower. One stage detectors, like YOLO, predict everything in a single pass, so they're fast enough for video. Today, modern one stage models are excellent, and they're the usual default.
CNN or vision transformer?
CNNs assume local patterns and work well with less data. Transformers see globally, and they need more data, or strong pretraining. On small datasets, a pretrained ConvNeXt is a great choice. With lots of data, or for rich tasks, transformers shine. Many modern models mix the two, and in practice, your data matters more than this choice.
Summary
Requirements, lighting, data, and deployment. Talk about those, and you'll sound like a professional. Next, your cheat sheet, and the road ahead.
Key takeaways- Start with requirements
- Name the trade-offs
- Talk about deployment