Computer Vision Engineer Interview Prep Guide

Cover parameter sharing (same filter across spatial locations), translation equivariance, hierarchical feature extraction from edges to objects, and dramatically reduced parameter count compared to fully connected layers. Explain stride, padding, receptive field growth through depth, and how 1x1 convolutions reduce channel dimensionality. Discuss why Vision Transformers are now competitive and when CNNs still have advantages.

Cover the full pipeline: data collection with dermatologist labeling and IRB approval, handling class imbalance (skin cancer is rare), data augmentation specific to medical imaging, model architecture selection, calibration for high sensitivity while managing false positive rate, uncertainty quantification, and regulatory requirements (FDA clearance). Discuss the clinical workflow: how model predictions integrate with dermatologist review, not replace it.

YOLO frames detection as a single regression problem, predicting bounding boxes and class probabilities in one forward pass. This gives speed but historically sacrificed accuracy for small objects. Faster R-CNN uses a region proposal network followed by a classification stage, giving better accuracy but slower inference. Discuss YOLOv8/v9 improvements that have narrowed the accuracy gap, anchor-free approaches, and when each is appropriate based on latency requirements.

Show your debugging methodology for ML projects: error analysis on failure cases, dataset investigation, architecture experiments, loss function modifications, and training strategy adjustments. Quantify improvements at each iteration. Demonstrate that you approach CV development scientifically with hypotheses, controlled experiments, and systematic evaluation rather than random architecture search.

Cover model optimization techniques: knowledge distillation, quantization (post-training and quantization-aware training), pruning, architecture search for mobile (MobileNet, EfficientNet-Lite), and export to optimized runtimes (TensorRT, Core ML, TFLite, ONNX). Discuss the accuracy-latency tradeoff, how to measure real-world performance on target hardware, and battery consumption considerations for mobile deployment.

Systematic investigation: compare production data distribution to training data (domain shift), analyze failure cases by category (lighting, occlusion, novel objects, image quality), check preprocessing pipeline differences between training and serving, verify model version and configuration deployed correctly, and examine edge cases the test set did not cover. Discuss building a production evaluation pipeline with continuous monitoring for accuracy degradation.

ViT splits images into fixed-size patches, projects them linearly, adds positional embeddings, and processes them through standard transformer encoder layers with self-attention. Unlike CNNs, ViT has no built-in inductive bias for translation equivariance or locality, requiring more data to learn these properties. Discuss when ViT outperforms CNNs (large datasets, large model sizes) and where CNNs remain competitive (small datasets, edge deployment).

Discuss data collection strategy, annotation tool selection, labeling guidelines creation, quality control processes (inter-annotator agreement, expert review), handling edge cases in annotation, and iterative refinement of the dataset based on model error analysis. Show that you understand data quality is often the bottleneck for model performance and that building good datasets is an engineering discipline, not just a labeling task.