Computer Vision interview questions

Compare Faster R-CNN and YOLO in terms of accuracy and speed.

When should data augmentation be used, and what types are most commonly applied?

Explain what Semantic Segmentation is and how U-Net implements this task.

What is quantization and how does it help in deploying models on resource-constrained devices?

How do contrastive learning methods like SimCLR work and what are they used for?

What is convolution using stride and padding parameters, and how do they affect the size of the output data?

How does the NMS (Non-Maximum Suppression) algorithm work in object detection tasks, and what problems does it solve?

How would you obtain a video embedding for a recommendation system? Which models and approaches would you use?

Describe in detail how you approached car color recognition: data collection, model, training, and aggregation.

How did you collect the dataset and train a metric learning model to find fake car entries in the database?

How would you move from 17 fixed colors to an arbitrary set of over 100 colors, including 'silver metallic'? Propose a pipeline.

How did you aggregate metric learning distances across different views? Why is boosting better than heuristics?

What color format does OpenCV use to store images by default?

Name at least three image interpolation methods.

For the car classification model, did you use a pretrained model or train from scratch?

What experience do you have with video computer vision tasks, such as tracking, segmentation, or action recognition?

How would you consistently segment objects that a person interacts with in a video?

How would you approach automatic pixelation or censoring of video content?

How would you adapt a model trained on first-person VR videos to ordinary third-person videos when data is scarce?

Describe the embedding model training in more detail: architecture, data, and metrics.