Question 1
Difficulty: medium
Tell me about a computer vision project where you had to improve model accuracy under real-world conditions.
Sample answer
In one project, I worked on an object detection system for warehouse operations, and the model performed well in the lab but dropped noticeably once it was deployed. The main issue was variability in lighting, camera angle, and motion blur from moving carts and people. I started by analyzing failure cases instead of just tuning hyperparameters blindly. That led me to improve the training set with harder examples, add stronger augmentations, and rebalance classes that were underrepresented. I also switched from a single validation split to a more realistic evaluation strategy that mirrored production conditions. After that, I fine-tuned the detector and measured performance by scene type, not just overall mAP. The result was a meaningful lift in robustness and fewer false negatives. What I learned was that in computer vision, better data often has more impact than a more complex model, especially when the deployment environment is messy and dynamic.
Question 2
Difficulty: medium
How do you decide whether to use classical computer vision techniques or a deep learning approach for a problem?
Sample answer
I usually start by understanding the problem constraints, not the algorithm. If the task has limited data, clear geometric rules, or needs very low latency on constrained hardware, classical methods can still be the right choice. For example, thresholding, contour detection, optical flow, or feature matching can be effective for stable environments and are easier to explain and deploy. If the visual patterns are complex, variable, or difficult to hand-engineer, then I lean toward deep learning. I also consider maintenance: a traditional pipeline may be faster to prototype, but a learned model can generalize better if the data keeps changing. In practice, I often use a hybrid approach. For instance, I might use classical preprocessing for image normalization or region extraction, then apply a neural network for classification or detection. My decision is guided by accuracy, latency, interpretability, and how much labeled data is realistically available.
Question 3
Difficulty: hard
Describe how you would handle a computer vision model that performs well in training but poorly in production.
Sample answer
When that happens, I assume there is a mismatch somewhere between the training setup and the real environment. My first step is to compare the data distributions: image quality, camera type, resolution, lighting, and class balance. I also look for label issues, because noisy annotations can make a model look better than it really is. Next, I inspect production failures manually and group them into patterns such as blur, occlusion, unusual viewpoints, or rare classes. That usually reveals whether the issue is data, architecture, preprocessing, or thresholding. From there, I may retrain with harder examples, improve augmentation, calibrate confidence scores, or adjust the post-processing logic. I also like to track performance by slice, not just overall metrics, because one problem area can be hidden in the average. My goal is to close the loop quickly: diagnose, fix, validate, and then monitor after deployment so the same failure mode does not come back.
Question 4
Difficulty: medium
What metrics do you use to evaluate an object detection model, and why?
Sample answer
I choose metrics based on the business goal, but for object detection I usually look at precision, recall, F1, and mAP as a starting point. mAP is useful because it summarizes performance across confidence thresholds and IoU levels, so it gives a broad view of detection quality. But I never rely on it alone. If missing an object is expensive, recall becomes especially important. If false alarms create a lot of manual work, precision matters more. I also pay attention to per-class performance, because average numbers can hide weak classes that are operationally important. In some cases, I evaluate localization quality separately, especially if the box accuracy affects downstream actions like cropping or tracking. For deployment, I often add latency and throughput as operational metrics, because a highly accurate model that cannot meet response-time requirements is not a viable solution. The best metric set is the one aligned with the actual product impact.
Question 5
Difficulty: medium
Tell me about a time you had to work with imperfect or limited labeled data.
Sample answer
I worked on a defect detection problem where labeled images were expensive because they required expert review. We had a small initial dataset and a lot of unlabeled images from the production line. Instead of waiting for a perfect dataset, I proposed a staged approach. First, we defined a strict labeling guideline so the annotations would be consistent. Then I used the limited labeled set to train a baseline model and identify uncertain samples. Those samples were prioritized for review, which made the labeling process more efficient. I also used augmentation carefully to expand the training diversity without creating unrealistic images. In parallel, I ran error analysis to find which defect types the model struggled with most. That helped us target the next round of annotation more intelligently. The result was not just a better model, but a better data collection process. In my experience, strong dataset strategy can matter as much as model architecture when labels are scarce.
Question 6
Difficulty: medium
How do you approach choosing and tuning a model architecture for a new vision task?
Sample answer
I begin with the simplest architecture that could plausibly solve the problem. That means I first define the task clearly: classification, detection, segmentation, tracking, or some combination. Then I look at constraints such as inference speed, memory limits, and how much labeled data I have. If I need a fast prototype, I often start with a proven baseline like a ResNet-based classifier or a standard detector such as YOLO or Faster R-CNN depending on the use case. I tune architecture only after I have a baseline and a clear understanding of failure cases. When tuning, I focus on what the data is telling me: do I need higher resolution, better multi-scale feature extraction, stronger backbones, or a lighter model for edge deployment? I prefer iterative improvement over big, untested changes. That keeps experiments interpretable and makes it easier to understand which improvement actually moved the needle. In vision, architectural complexity should solve a real problem, not just look impressive.
Question 7
Difficulty: hard
How would you optimize a vision model for edge deployment or real-time inference?
Sample answer
I optimize for edge deployment by treating accuracy and efficiency as equal priorities from the start. First, I understand the hardware target, because what works on a server may not work on a mobile device or embedded GPU. Then I look at the full pipeline, not just the network: image preprocessing, inference, and post-processing can all affect latency. Common steps include reducing input resolution if the task allows it, choosing a lighter architecture, and applying quantization or pruning where appropriate. I also benchmark different runtime frameworks because implementation details can make a big difference. If accuracy drops after compression, I check whether the model is sensitive to quantization and whether calibration data is representative. In some projects, distillation has been useful for keeping a smaller model close to the performance of a larger one. I always validate with real latency measurements on the target device, because theoretical speedups do not always hold in production. Practical deployment is about tradeoffs, not perfection.
Question 8
Difficulty: medium
Describe a situation where you disagreed with a teammate or stakeholder about a model approach. How did you handle it?
Sample answer
In one project, a stakeholder wanted to push for a very complex model because they believed higher complexity would automatically mean better performance. I understood the concern, but the available data was small and the deployment environment was constrained, so I thought the risk was high. Instead of just saying no, I presented a comparison plan with three baselines: a simple model, a medium-complexity model, and the proposed heavier one. I also defined success criteria beyond accuracy, including latency, stability, and ease of maintenance. Once we ran the experiments, the simpler or mid-sized models actually performed nearly as well, and they were much easier to deploy. That made the discussion objective rather than personal. I try to handle disagreements by bringing evidence, not ego. In computer vision, there are often multiple technically reasonable paths, so the best way to align people is to make the tradeoffs visible and let the data support the final choice.
Question 9
Difficulty: hard
How do you debug a segmentation model that looks accurate overall but fails on specific boundaries or small objects?
Sample answer
When a segmentation model has strong overall metrics but poor boundaries or small-object performance, I dig into the evaluation rather than trusting the average score. I look at class-wise IoU, boundary quality, and performance by object size. Often the issue comes from insufficient resolution, coarse labels, or a model that is too focused on global context. I would inspect predictions visually on hard examples and compare them to the annotations to see whether the problem is systematic. If small objects are being missed, I may increase the input size, adjust the feature pyramid, or modify the loss to better emphasize rare regions. For boundary issues, I might add boundary-aware losses or refine post-processing. I also review the annotations themselves, because inconsistent masks can make a model appear worse than it is. In practice, segmentation debugging is a mix of metrics, visual inspection, and dataset quality checks. The key is to understand what the model is actually learning, not just whether the headline number looks acceptable.
Question 10
Difficulty: easy
Why do you want to work as a Computer Vision Engineer, and what strengths would you bring to the role?
Sample answer
I enjoy computer vision because it sits at the intersection of data, software, and real-world impact. I like that the work is never just about training a model; it also involves understanding the product, the environment, and the operational constraints. That combination keeps the role challenging in a good way. My strongest contribution is that I’m disciplined about turning ambiguous visual problems into measurable experiments. I’m comfortable with data analysis, model training, debugging, and deployment considerations, so I can move from idea to implementation without losing sight of the end goal. I also pay close attention to failure analysis, because that is usually where the biggest improvements come from. In team settings, I communicate technical tradeoffs clearly and focus on practical outcomes. I think a strong computer vision engineer has to be both analytical and adaptable, and that is the type of work I do best. I like building systems that keep improving after they leave the lab.
