Computer Vision Engineer Interview Prep: Questions and a Mock Test

Expect to be asked why a convolutional architecture suits images rather than a fully connected network, and the answer should be about structure rather than about performance.

Three properties do the work. Local connectivity, because a pixel's meaning depends mostly on its neighbours, so each unit looks at a small region rather than the whole image. Parameter sharing, because a feature such as an edge is worth detecting anywhere, so the same filter slides across the image; this is what makes the parameter count independent of image size. And translation equivariance, because shifting the input shifts the feature map correspondingly, so the network does not need to learn the same object separately at every position.

The practical vocabulary follows. A receptive field is the region of the input that influences one output unit, and it grows with depth, with larger kernels, and with stride and pooling. Stacking two three-by-three convolutions gives the same receptive field as one five-by-five with fewer parameters and an extra non-linearity, which is the reasoning behind small-kernel architectures. Dilated convolutions expand the receptive field without adding parameters or reducing resolution, which is why they appear in segmentation.

Residual connections are worth being able to justify: adding the input to the output gives gradients a path that skips the block, which is what made very deep networks trainable. Batch normalisation stabilises training by normalising activations, and the detail interviewers probe is that it behaves differently at training and inference time, using batch statistics during training and running estimates afterwards.

Being precise about the task is the first thing assessed, because the metrics and the failure modes differ.

Classification assigns a label to a whole image. Object detection localises instances with boxes and labels. Semantic segmentation labels every pixel with a class but does not separate touching instances of it. Instance segmentation separates them. Panoptic segmentation combines instance-level things with amorphous stuff such as sky and road.

The design distinction most often asked about is one-stage versus two-stage detection. Two-stage approaches propose candidate regions and then classify and refine them, which historically gave better accuracy at higher latency. One-stage approaches predict boxes and classes directly in a single pass and are faster, with the accuracy gap having narrowed substantially. Transformer-based detectors remove hand-designed components such as anchor boxes and non-maximum suppression by treating detection as set prediction.

Non-maximum suppression is a favourite question because it is a post-processing step with real consequences. A detector emits many overlapping boxes for one object, and suppression keeps the highest-scoring box and removes others overlapping it above a threshold. The tradeoff to name: a low threshold removes duplicates aggressively and also removes genuine detections of objects that genuinely overlap, such as people in a crowd. That is a tuning decision driven by the deployment, not a constant.

Intersection over union is the underlying measure, the area of overlap divided by the area of union of two boxes, ranging from zero to one.

Metric questions separate candidates quickly, because the standard reporting hides specific failures.

Start from the definitions. scikit-learn states them exactly: precision is "the ability of the classifier not to label as positive a sample that is negative", computed as true positives over true positives plus false positives; recall is "the ability of the classifier to find all the positive samples", true positives over true positives plus false negatives.

Average precision summarises the precision-recall curve, defined as the sum over thresholds of the change in recall multiplied by the precision at that threshold. One detail is worth carrying into an interview: the documentation warns that linear interpolation of points on the precision-recall curve "provides an overly-optimistic measure of classifier performance", which is what the trapezoidal rule does when computing area under that curve. So average precision computed by interpolation and average precision computed without it are not the same number, and comparing across implementations that differ on this is comparing incomparable figures.

For detection, mean average precision aggregates average precision across classes, and modern benchmarks average across several intersection over union thresholds rather than a single loose one, which rewards precise localisation rather than merely finding the object somewhere.

The judgement half is that a single aggregate hides everything. A detector can improve overall while getting worse on small objects, on a particular camera, at night, or on the rarest and most important class. Per-slice evaluation is what catches that, and proposing it unprompted is a strong signal. Accuracy on an imbalanced problem is close to meaningless, since predicting the majority class scores well while finding nothing.

In applied vision work the dataset is more often the bottleneck than the architecture, and interviewers with production experience ask about it directly.

Label quality comes first. Bounding box conventions vary, tight versus loose, whether occluded parts are included, how truncated objects at the frame edge are handled, and inconsistency between annotators produces a ceiling on measurable accuracy that no model can exceed. Measuring inter-annotator agreement, and writing a labelling guide before annotation rather than after, are the answers that indicate experience.

Class imbalance is second, and vision datasets are almost always imbalanced. The available responses each have a cost: oversampling rare classes risks overfitting to the few examples, undersampling discards data, class-weighted loss is simple and can destabilise training, and focal loss reduces the contribution of easy examples so training focuses on hard ones.

Augmentation is third, and the question interviewers use is which augmentations are valid for the task. Horizontal flips are fine for most natural images and wrong for text or for anything where handedness matters. Rotation is fine for aerial or medical imagery and can be wrong for street scenes. Colour jitter helps when lighting varies in deployment and hurts when colour is the discriminating feature. The principle to state: augmentation should express variation you expect at inference, not arbitrary distortion.

Google's Rules of Machine Learning applies here too, particularly the advice to "keep the first model simple and get the infrastructure right", because "the first model provides the biggest boost to your product, so it doesn't need to be fancy".

Deployment rounds ask about the constraints that decide architecture choices in practice.

The latency budget usually determines the model more than accuracy does. A model that takes 200 milliseconds cannot serve a 30 frames per second video stream on one device, and knowing the levers is expected: a smaller backbone, reduced input resolution, quantisation to lower precision, pruning, and batching where the workload allows it. Quantisation is the one most often discussed, and the useful distinction is between post-training quantisation, which is quick and can lose accuracy, and quantisation-aware training, which costs a training run and usually recovers most of it.

Edge deployment adds constraints that cloud deployment does not: fixed memory, no ability to scale out, thermal throttling under sustained load, and a model that must be updated over a network you do not control.

Monitoring is where vision systems fail quietly. Labels rarely arrive in production, so the practical signals are the input distribution, image statistics such as brightness and blur, and the prediction distribution, including how often the model abstains or returns low confidence. A camera that is knocked out of position, a lens that fouls, or a seasonal lighting change will degrade a model without any code changing. Google Cloud's MLOps framing is the right reference for the automation around this, where level 1 requires automated data and model validation rather than only a retraining schedule.

One closing point interviewers like: a confidence threshold is a product decision. Where you set it determines whether the system annoys people with false alarms or misses what it was installed to catch.