21 Computer Vision Machine Learning Interview Questions

Interviews for computer vision engineer and CV researcher roles are similar in a lot of ways to ML engineer interviews. You’ll face lots of ML system design and Python programming questions.

However, the biggest difference is that all the questions will relate to computer vision, and during the technical screens and the on-site, you’ll be facing questions and sessions that are focused on CV techniques.

The most common concepts to come up in computer vision interviews include:

• Computer vision algorithms and ML - designed to assess your knowledge of computer vision algorithms, like convolutional neural networks, as well as CV-related machine learning concepts.
• Programming - expect a technical screen covering Python, as well as on-site sessions focused on Python CV techniques
• Computer vision basics - consists of defining particular techniques, CV libraries, and more.

To help, we’ve compiled some top computer vision interview questions.

Early on in the interview process, be prepared for basic computer vision interview questions. These questions are designed to quickly assess your expertise and working knowledge of common CV tools.

One tip: Always check out the computer vision projects that a company is working on before getting started. You can find CV blogs for most large tech companies – here are the blogs from Facebook, Amazon, and Google.

1. What is a computer vision library? What are the most common CV libraries?

Interviewers ask questions like this to quickly understand your expertise and working knowledge of CV tools. A computer vision library is used to store equations and functions that can then be used to create neural networks for image identification and processing. Popular CV libraries include:

• TensorFlow
• OpenCV
• MATLAB
• Keras
• PCL
• DeepFace

2. What technique(s) would you use to evaluate an object detection model?

As you prepare, practice defining use cases that align with specific CV techniques and potential benefits. For this question, you might discuss Intersection over Union (IoU). IoU is a metric used to evaluate the accuracy of the model and measures the amount of overlap between the predicted and ground truth bounding box.

3. What is OpenCV? What algorithms are available in OpenCV?

OpenCV is a computer vision library that offers a variety of computer vision programming functions. It’s mostly written in C++, but there is a JavaScript version, as well. Some of the most common algorithms that can be deployed in OpenCV are random forest, artificial neural network, support vector machine, convolution neural networks.

4. Explain the “mach band effect.”

This refers to the optical illusion that is created when images have edges with similar shades of grey. The eye naturally adjusts by interpreting higher contrast between the two. In computer vision, the mach band causes inaccurate results. One way to reduce the mach band effect is to adjust the smoothness to reduce the banding effect.

Additional CV Basics Questions

• Explain why the inputs in computer vision programs can get so large. What are some methods for overcoming this?

• Why would you greyscale images in computer vision? Explain what a digital image is?

• How does convolution work? How does it change if the inputs are greyscale vs. RGB?

Machine learning algorithm questions in computer vision interviews assess your understanding of CV algorithms, e.g. how they work, use cases, and how you might go about implementing a new model. You should have a strong grasp of SIFT, Global Block Matching, and convolutional neural networks, as well as traditional machine learning algorithms.

1. What does it mean to use “inception architecture” for a CNN? What does it solve?

An inception network is a deep neural network with an architecture that consists of inception modules, which are repeating components. Using inception architecture introduces inception blocks, which contain multiple convolutional and pooling layers stacked together. This gives more accurate results and can help to reduce computation costs.

2. How would you encode a categorical variable with thousands of distinct values?

The approach to this encoding categorical features question depends on whether the problem is a regression or a classification model. If it’s a regression model, one possible solution would be to cluster them based on the response by working backwards. You could sort them by the response variable and then split the categorical variables into buckets based on the grouping of the response variable.

3. How would you explain bias variance tradeoff in machine learning to a high school student?

Bias variance tradeoff is the tradeoff between predictive accuracy and generalization of patterns outside training data. Increasing the accuracy of the model will lead to less generalization, while increasing the bias will decrease the variance.

4. How would you design the inputs and outputs for a model to detect bombs at a border crossing. Measure accuracy? Test your model?

Remember the equation for precision:

Because we can’t have high TrueNegatives, recall should be high when assessing the model.

Additional Computer Vision Algorithm Questions

• How would you modify a neural network to account for mislabeled training data?
• How would you explain the predictions of a neural network to non-technical stakeholders?

Python machine learning questions test your ability to write code that’s used in computer vision. This would include writing functions for image rotation, image processing, detection and visualization.

1. Given an array filled with random values, write a function rotate_matrix to rotate the array by 90 degrees in the clockwise direction.

Is there a simpler matrix transformation that can be performed on the example matrix array that gives the same output as a 90 degree clockwise rotation? What is special about the example matrix that makes this possible?

2. Given a JSON object with nested objects, write a function flatten_json that flattens all the objects to a single key-value dictionary.

You’ll likely be asked to perform the task without using a library.

3. Build a KNN classification model from scratch.

Writing algorithms from scratch are increasingly common in coding interviews. With a KNN algorithm question like this, you’ll be given conditions, and then will be instructed to write the code from the ground up.

Prepare for your computer vision and machine learning interviews with these resources from Interview Query:

• Machine Learning Interview Questions
• Machine Learning Algorithm Questions
• Interview Query’s Machine Learning Course