Computer Vision Internship Interview Questions for Freshers (with Answers)
By K. Boopathikumar, CEO & Founder · 16 November 2026 · 8 min read
Computer vision interns build systems that understand images and video. Interviews test image basics, CNNs, the difference between vision tasks, and how you measure a model.
Image basics
- How is an image stored? As an array of pixel values: height x width x channels (3 for RGB). OpenCV loads colour images in BGR order, a common source of bugs.
- What is convolution? Sliding a small filter over the image and computing weighted sums, which detects patterns such as edges.
- Why normalise images? Scaling pixel values (for example to 0-1 or by dataset mean and standard deviation) makes training stable and matches pretrained models.
- What does pooling do? Reduces the spatial size of feature maps, keeping the strongest signals and adding some position tolerance.
Models and tasks
- Classification vs detection vs segmentation? Classification labels the whole image; detection draws boxes around objects; segmentation labels every pixel.
- What is transfer learning? Starting from a model pretrained on a large dataset and fine-tuning it on your smaller one. It saves data and time.
- Why use data augmentation? Flips, crops, rotations and colour changes create varied examples, reducing overfitting. Only use changes that make sense, such as no vertical flips for road scenes.
- What is YOLO? A family of fast one-stage object detectors that predict boxes and classes in one pass, popular for real-time use.
Evaluation and deployment
- What is IoU? Intersection over union: the overlap between a predicted box and the true box divided by their combined area.
- What is mAP? Mean average precision across classes and IoU thresholds, the standard metric for detection.
- Why does a model fail in the real world after good test scores? Different lighting, cameras, angles or backgrounds from the training data (domain shift). Collect data from the real setting.
- How do you run a vision model on a phone or edge device? Use a small architecture, quantise it and export to formats such as ONNX, TensorFlow Lite or OpenVINO.
- What privacy issues come with vision? Faces and number plates are personal data; collect with consent, blur where possible and store securely.
How to answer well in the interview
- Show sample predictions, including failures, not only metrics.
- Explain how you labelled data and checked label quality.
- Mention inference speed in frames per second if real-time matters.
- Discuss privacy and consent for any images of people.
Next steps
- Study the free Computer Vision learning path and take its module quizzes for a free certificate.
- Follow the Computer Vision skill track from basic to advanced.
- Do real, reviewed work in a Computer Vision internship and keep every project in your own GitHub.