Computer Vision Engineer Interview Questions for AI Training Work

I use a single-stage detector when inference speed matters more, like a real-time application, since two-stage detectors are generally more accurate but slower. If the application can tolerate more latency and accuracy is the priority, like offline batch processing, a two-stage approach is often worth the extra cost.

I use a combination of class-weighted loss functions and targeted data augmentation for underrepresented classes, rather than relying on oversampling alone, since naive oversampling can lead to overfitting on the minority class. I also check whether the imbalance reflects real-world distribution, since forcing artificial balance can hurt real-world performance.

I test on a held-out set that deliberately varies those conditions rather than just a random split of the same distribution, since a random split can share the same biases as training data. If performance drops significantly under different conditions, that's a signal the model learned spurious correlations rather than the actual task.

I match preprocessing to what the model architecture expects and what the deployment environment will actually produce, rather than applying a standard pipeline by default. Normalization and resizing are almost always necessary, but more aggressive preprocessing, like heavy denoising, depends on the actual noise characteristics of the input data.

I look at quantization and pruning first, since they often reduce model size and latency significantly with a manageable accuracy tradeoff, before considering a smaller architecture from scratch. I benchmark on the actual target hardware rather than assuming desktop GPU performance translates to the edge device.

I'd compare the distribution of production images against the training data directly, since real-world images often differ in lighting, resolution, or occlusion patterns that weren't well represented in training. I'd collect a sample of failure cases and look for a pattern rather than assuming the failures are random.

I'd look at transfer learning from a model pretrained on a large, related dataset first, fine-tuning only on the limited labeled data available, rather than training from scratch. I'd also consider whether synthetic data generation or targeted data augmentation can meaningfully expand the effective training set.

I'd evaluate on data that closely mirrors actual deployment conditions rather than relying solely on a standard benchmark, and I'd look at failure cases qualitatively, not just aggregate accuracy numbers. A model can hit strong benchmark numbers while still failing in ways that matter for the specific use case.

A detection model performed noticeably worse on images taken in low-light conditions, which correlated with a specific deployment environment. I traced it to underrepresentation of that condition in training data and added targeted data collection and augmentation for low-light scenarios, which closed most of the gap.

A real-time application needed to run on modest hardware, and the most accurate model I'd built was too slow to meet the latency requirement. I evaluated several smaller architectures and ultimately chose one with a small accuracy tradeoff that met the speed requirement, since a highly accurate model that couldn't run in real time wasn't usable at all.