What should you focus on for a computer vision interview?
Which model questions should you check first?
Start with how models work, not just their names. The CNN architecture question appeared in 5 of 75 interviews, making it the most asked question in these interviews. Other repeated questions asked about convolution, detector choices, and non-maximum suppression.
- Can you explain what convolution does and what convolutional layers contribute to a CNN?
- Can you compare YOLO with two-stage detectors such as Faster R-CNN and say when you would choose each?
- Can you explain non-maximum suppression and why it is needed in object detection?
Use these as separate checks. Explaining a CNN does not cover the detector comparison or the reason for non-maximum suppression.
How much attention should you give data and annotation?
Data and annotation appeared in 21 of 75 interviews. Give labeling decisions attention alongside model choices. This topic covers collecting, storing, preparing, and labeling data, as well as checking its quality.
- Can you explain how you would measure and improve agreement between annotators on subjective video labels?
- Can you set labeling rules for objects hidden from view or partly outside the image?
- Can you say when a tracked object should keep its ID and when that ID should end?
- Can you describe how you would check that labels stay consistent across video frames?
Your checks should cover both the labeling rules and how you would review the finished labels.
Can you explain how you would test a system?
Face recognition questions include evaluation tasks, data, and metrics. The training and evaluation topic also covers dataset splits, leakage, losses, regularization, and model failures.
- For face recognition, can you choose evaluation tasks, test data, and metrics? Can you explain enrollment, matching thresholds, and the choice of reference images?
- For OCR, can you describe how you would validate extracted text?
- For model training, can you explain how you would split the data and check for leakage?
Can you explain application and deployment choices?
OCR and vision-language models appeared in 17 of 75 interviews. Check whether you can choose between them based on accuracy, cost, and latency, then explain where OCR fits into an application.
For deployment, can you explain how you would measure latency and throughput for a camera stream? Can you compare running the model on a device with running it in the cloud?
For system design, can you name a trade-off in a vision project and explain how that choice affected the result?
Who gets asked Computer Vision
The 20 most asked Computer Vision questions
Counted by interview, not by mention: a question asked twice in one conversation counts once.
How does a CNN architecture work, and what role do convolutional layers play?
Asked in 5 of 75 interviews · last asked Sep 23 · Concept · Junior
Show answer
A CNN learns image features through stacked layers, then uses them for tasks such as classification. Convolutional layers slide learned filters across the image or feature maps to detect local patterns. Early layers learn edges; deeper layers combine them into more complex features. Activation functions add nonlinearity, while pooling or strided convolutions can reduce spatial size.
- Each filter shares its weights across spatial positions, reducing parameters compared with fully connected layers.
- Convolution produces feature maps that show where learned patterns appear.
- Deeper layers usually combine information from larger parts of the input.
- The output head depends on the task; classification often uses global pooling followed by a linear layer.
What the interviewer checks: The interviewer checks whether you understand how convolutions extract features and how CNN layers work together.
Other ways it was asked
- “Can you explain CNN architecture?”
- “Can you explain the CNN architecture?”
- “What is a convolution in a convolutional neural network, and how does the architecture work?”
How does YOLO compare with two-stage detectors such as Faster R-CNN, and when would you choose each?
Asked in 3 of 75 interviews · last asked Aug 27 · Concept · Mid-level
Show answer
YOLO predicts boxes and classes directly, while Faster R-CNN first proposes regions, then classifies them and refines their boxes. I would choose YOLO for low-latency video or edge devices. I would consider Faster R-CNN when accuracy matters more than speed, but compare both on the target data and hardware: newer YOLO models can be highly accurate.
- YOLO is a one-stage detector; Faster R-CNN is a two-stage detector.
- Faster R-CNN shares image features across both stages, but region processing adds work.
- Speed and accuracy depend on model size, input resolution, hardware, and data.
- Test small-object accuracy and crowded scenes rather than assuming either detector always wins.
What the interviewer checks: The interviewer checks whether you understand the architectures and can choose based on measured speed and accuracy needs.
Other ways it was asked
- “What is the difference between YOLO and Faster R-CNN?”
- “What is a use case for YOLO, and under what conditions might it be preferable to Faster R-CNN?”
- “Can you tell me the difference between YOLO and a two-stage detector?”
What evaluation tasks, data, and metrics would you use to measure face recognition accuracy?
Asked in 3 of 75 interviews · last asked Oct 11 · Scenario · Mid-level
Show answer
I would measure one-to-one verification and one-to-many identification separately. I would use labeled face images with different lighting, poses, cameras, and demographic groups, keeping test identities out of training. For verification, I would report false accept and false reject rates at chosen thresholds; for identification, I would report rank-one accuracy and errors when rejecting unknown people.
- Verification needs both same-person and different-person pairs; overall accuracy alone can hide false accepts.
- Choose similarity thresholds on validation data, then keep them fixed during testing.
- Identification tests should match the real gallery size and include people not enrolled.
- Report results by demographic group and image condition; avoid duplicate images across data splits.
What the interviewer checks: The interviewer checks whether you can design realistic tests and explain the trade-off between false matches and missed matches.
Other ways it was asked
- “How should you measure and interpret the accuracy of a face-recognition system such as DeepFace?”
- “What data or evaluation task did you use to measure a face recognition model's accuracy?”
- “Which features and evaluation metrics would you use to assess face matching accuracy in an attendance system?”
How does convolution work in a neural network?
Asked in 2 of 75 interviews · last asked Sep 23 · Concept · Mid-level
Show answer
A convolutional layer slides learned filters over an input to detect local patterns. At each position, it multiplies filter weights by matching input values, sums the results, and usually adds a bias. Each filter produces one feature map. The same filter weights are reused at every position, reducing the number of parameters.
- In a standard convolution, each filter covers a small spatial area and all input channels.
- Stride controls how far the filter moves; padding controls how input borders are handled.
- Training learns the filter weights through backpropagation.
- An activation function, such as ReLU, often follows convolution to add nonlinearity.
What the interviewer checks: The interviewer checks whether you understand local filtering, shared weights, and how feature maps are produced.
Other ways it was asked
- “What is convolution, and how is it used in neural networks?”
- “What is convolution, and how does it work in a neural network?”
- “Can you explain how a convolutional layer works in deep learning?”
How would you measure and improve inter-annotator agreement for subjective video segmentation labels?
Asked in 2 of 75 interviews · last asked Sep 11 · Scenario · Senior
Show answer
I would have annotators label the same representative clips independently, then measure agreement by label type. I would use mask IoU for spatial regions and weighted kappa for ordered labels like occlusion severity. I would set review thresholds using a pilot and task needs. To improve agreement, I would clarify guidelines, review disputes, and repeat independent labeling.
- Define the labeling unit, severity levels, and acceptable temporal boundary differences before measuring agreement.
- Check agreement by class and difficult case; one overall score can hide problems.
- Use disputed examples to improve guidelines, with an expert resolving remaining disagreements.
- Allow uncertain labels when evidence is unclear; agreement alone does not prove correctness.
What the interviewer checks: The interviewer checks whether you can measure label consistency fairly and reduce ambiguity without forcing false agreement.
Other ways it was asked
- “In a video segmentation project, how would you define and measure inter-annotator agreement?”
- “In a video segmentation project, how would you measure inter-annotator agreement, set thresholds, and handle disputes during review?”
- “How do you measure and improve inter-annotator agreement when the task is subjective, such as occlusion severity?”
What tracking annotation rules would you set for occlusion, truncation, and retaining or ending object IDs?
Asked in 2 of 75 interviews · last asked Sep 11 · Scenario · Senior
Show answer
Keep the same ID through short occlusions when the object's identity is clear. Mark occlusion for objects hidden by other objects, and truncation for objects cut off by the frame. Start a track when an object is identifiable; end it after confirmed exit or an unresolved gap. Define whether boxes cover visible pixels or the estimated full object.
- Set a clear gap limit and re-entry rule; there is no universal policy.
- For fully hidden objects, define whether to skip boxes or estimate them with an explicit flag.
- Keep partially visible objects on the same track when identity is clear; never reuse IDs for different objects.
- Use labeled examples and review uncertain cases to keep annotators consistent.
What the interviewer checks: The interviewer is checking whether you can make tracking labels consistent without guessing object identity or visibility.
Other ways it was asked
- “When annotating videos for object tracking, what rules would you set for when to start, stop, or continue a track through short occlusions?”
- “How do you define and enforce annotation guidelines for occlusion and truncation in a track?”
- “How do you define and enforce tracking annotation guidelines for retaining or ending an object ID and labeling partially visible objects?”
What is non-maximum suppression, and why is it needed in object detection?
Asked in 2 of 75 interviews · last asked Sep 23 · Concept · Mid-level
Show answer
Non-maximum suppression (NMS) removes duplicate detections of the same object. It keeps the highest-scoring box, removes lower-scoring boxes whose overlap with it exceeds a threshold, and repeats with the remaining boxes. It is needed because detectors often predict several boxes around one object.
- Overlap is usually measured by intersection over union (IoU).
- NMS usually runs separately for each class.
- A lower IoU threshold removes more boxes but may remove detections of nearby objects.
What the interviewer checks: The interviewer checks whether you understand duplicate detections, score-based selection, and the overlap threshold trade-off.
Other ways it was asked
- “What is non-max suppression, and why is it needed in YOLO?”
- “Can you explain non-maximum suppression (NMS)?”
How many reference images per person should a face recognition system use for reliable matching?
Asked in 2 of 75 interviews · last asked Oct 11 · Scenario · Mid-level
Show answer
There is no fixed number of reference images per person that guarantees reliable matching. One clear image can work, but several varied images often help with changes in pose, lighting, and appearance. I would choose the number by testing on real operating conditions, checking both false matches and missed matches.
- Image quality and useful variation matter more than the raw image count.
- Systems usually compare face embeddings, not raw images.
- Multiple images can produce separate templates or one combined template.
- More templates may increase false matches, so test the matching threshold.
What the interviewer checks: The interviewer is checking whether you understand reference image quality, variation, and how to validate matching reliability.
Other ways it was asked
- “Should a live face recognition system compare each face against only one stored image?”
- “How many images of each person does a face recognition system need for reliable identification?”
How is OCR integration used in an application?
Asked in 2 of 75 interviews · last asked Oct 23 · Tooling · Mid-level
Show answer
OCR integration lets an application read text from images or scanned documents. The app sends an image to an OCR library or service, then uses the returned text. It can make documents searchable, fill form fields, or extract receipt details. Important results should be checked because OCR can misread text.
- Improve image quality with rotation correction, cropping, or noise removal when needed.
- Choose an OCR tool that supports the required languages and document types.
- OCR reads text; extra rules or models may be needed to identify fields.
- Check uncertain results and protect sensitive documents when using external services.
What the interviewer checks: The interviewer is checking whether you can connect OCR to an application workflow and handle errors and privacy.
Other ways it was asked
- “What is OCR integration?”
- “How would you use OCR in an application?”
When would you choose OCR over a vision-language model, considering accuracy, cost, and latency?
Asked in 2 of 75 interviews · last asked Oct 21 · Concept · Mid-level
Show answer
I would choose OCR for reading text from clear documents, especially when cost and latency matter. OCR often gives more exact text with less compute than a large vision-language model. I would use a vision-language model when extraction needs visual context or complex layout understanding, but check for invented values. I would compare both on real samples.
- Accuracy depends on image quality, language, layout, and the specific models.
- OCR reads text; extracting fields may need extra layout rules or parsing.
- Vision-language models can interpret context, but may invent or change text.
- Measure field accuracy, total pipeline cost, and end-to-end latency on representative images.
What the interviewer checks: The interviewer is checking whether you match the tool to the task and test accuracy, cost, and latency.
Other ways it was asked
- “For extracting data from an image, would you prefer OCR or a vision model, and how would that affect accuracy, cost, and latency?”
- “When would you use OCR versus a vision-language model?”
How would you connect to an RTSP camera stream and process its video?
Asked in 2 of 75 interviews · last asked Sep 22 · Tooling · Mid-level
Show answer
I would open the camera’s RTSP URL with OpenCV, using an FFmpeg or GStreamer backend that supports RTSP. I would read decoded frames, resize and normalize them as the model expects, then run inference. I would check connection and frame-read errors, reconnect after failures, and limit buffering to avoid processing old frames.
- Check the camera URL, credentials, network access, and video codec support.
- Keep capture separate from slow inference, using a bounded queue and dropping old frames when needed.
- TCP retransmits lost packets but can add delay; UDP does not retransmit lost packets.
- Release the stream connection when stopping.
What the interviewer checks: The interviewer checks whether you can build a reliable video pipeline while keeping processing delay low.
Other ways it was asked
- “How do you process an RTSP video stream?”
- “How would you connect a computer vision application to an RTSP stream from a camera?”
What is SLAM?
Asked in 2 of 75 interviews · last asked Sep 24 · Concept · Junior
Show answer
SLAM means Simultaneous Localization and Mapping. It lets a system build a map of an unknown area while estimating its own position and orientation in that map. An autonomous robot or drone uses sensors, such as cameras or lidar, for SLAM, then uses the results to plan paths and navigate.
- Mapping and localization happen together because each depends on the other.
- Visual SLAM uses cameras; other SLAM methods use lidar or combine sensors.
- Recognizing a previously visited place can help correct accumulated position errors.
What the interviewer checks: The interviewer is checking whether you understand how SLAM supports navigation without a known map.
Other ways it was asked
- “What is SLAM, and how is it used in an autonomous unmanned system?”
How can images be stitched together, and what image-stitching functions or techniques can be used?
Asked once · last asked Mar 30 · Concept · Mid-level
Show answer
Stitch images by finding matching points in overlapping areas, estimating their alignment, and warping them into a shared view. Then blend the overlaps to hide seams and brightness differences. OpenCV offers cv2.Stitcher_create() with its stitch() method for a full pipeline, or separate functions for a custom pipeline.
- Use SIFT or ORB features and descriptor matching to find corresponding points.
- For planar scenes, use cv2.findHomography() with RANSAC, then cv2.warpPerspective() to align images.
- Use exposure correction, seam selection, and feathering or multiband blending to reduce visible joins.
- Parallax and moving objects can cause ghosting; one homography may not align the whole scene.
What the interviewer checks: The interviewer checks whether you understand the stitching pipeline, useful OpenCV tools, and common alignment problems.
What sensing components does LiDAR use to measure its surroundings, and how does that differ from a camera?
Asked once · last asked Mar 30 · Concept · Mid-level
Show answer
LiDAR uses a laser emitter and a light detector, usually measuring how long reflected pulses take to return. A camera uses a lens and an image sensor to record incoming light. LiDAR measures distance directly, while a standard camera captures a 2D image.
- LiDAR sends its own light; a standard camera relies on light from the scene.
- LiDAR measures distances in many directions to build a 3D point cloud.
- A standard camera records appearance, such as color and texture; depth requires extra processing or hardware.
What the interviewer checks: The interviewer checks whether you understand active distance sensing versus image capture.
How can you improve image quality so that converting a 14-bit image to 8-bit preserves as much of its quality as possible?
Asked once · last asked Apr 7 · Concept · Mid-level
Show answer
Map the useful intensity range to 0–255, using a tone curve when needed, then round to 8-bit. Choose the range carefully to avoid clipping important highlights or shadows. Add small dithering noise before rounding to reduce visible banding, but remember that 8-bit cannot retain all 14-bit detail.
- Simple bit shifting can waste contrast when the image uses only part of the 14-bit range.
- Gamma correction or tone mapping can preserve visible detail better than linear scaling.
- Stretching contrast improves visibility, but clipping extreme values loses information.
- Keep high precision during processing; quantize to 8-bit only at the end.
What the interviewer checks: The interviewer checks whether you understand intensity mapping, quantization loss, and ways to preserve visible detail.
How can transformers be applied to image data?
Asked once · last asked Jul 16 · Concept · Mid-level
Show answer
Transformers can process an image as a sequence of patches. Each patch becomes a vector, with position information added to show where it belongs. Self-attention lets patches share information across the image. The output can support tasks such as image classification, object detection, and segmentation.
- A Vision Transformer splits images into patches and maps each patch to an embedding.
- Position information helps the model understand the spatial layout.
- Full self-attention becomes costly as the number of patches grows.
- Local attention windows can reduce computation for large images.
What the interviewer checks: The interviewer is checking whether you understand how image patches become transformer inputs and the cost of attention.
How would you handle a super-resolution pipeline to enhance image quality?
Asked once · last asked Aug 3 · Scenario · Senior
Show answer
I would choose a super-resolution model based on the image type, noise, blur, and target scale. I would use a pretrained model, then fine-tune it on relevant image pairs if needed. I would compare results with bicubic resizing using quality metrics and visual checks. Before deployment, I would test speed, memory use, and unwanted artifacts.
- Training inputs should reflect real noise, blur, and compression, not just clean downsampling.
- Use PSNR and SSIM when reference images exist; also check visual quality.
- Sharper output can contain invented details, which is risky for medical images or evidence.
- Process large images in overlapping tiles when memory is limited, and check for visible seams.
What the interviewer checks: The interviewer is checking whether you can build and evaluate a practical pipeline while balancing quality, reliability, and speed.
Can you compare human visual perception with computer vision systems?
Asked once · last asked Aug 12 · Concept · Mid-level
Show answer
Humans use visual input, context, and past experience to understand a scene, while computer vision systems use algorithms and learned patterns. Humans usually adapt better to unfamiliar situations and need fewer examples. Computer vision can process many images quickly and perform well on narrow tasks, but changes in lighting or viewpoint can reduce accuracy.
- Humans use broad world knowledge; computer vision systems often rely on patterns learned from training data.
- Computer vision can outperform humans on specific tasks, but this does not mean it understands every scene.
- Both can make mistakes: humans face visual illusions, while models can fail on unfamiliar inputs.
What the interviewer checks: The interviewer checks whether you understand the strengths, limits, and differences between human and machine vision.
Does high accuracy in an image classification system mean that it has human-like visual perception?
Asked once · last asked Aug 12 · Concept · Mid-level
Show answer
No. High accuracy shows that a system predicts labels well on the tested images, not that it sees like a human. It may rely on background clues or textures instead of the object itself. Testing on unfamiliar images and misleading visual clues helps reveal these limits.
- Accuracy measures task performance, not how the system reaches its answer.
- Backgrounds, textures, or dataset bias can make classification easier without deeper visual understanding.
- Strong results on one dataset may not hold under changes in lighting, viewpoint, or setting.
What the interviewer checks: The interviewer is checking whether you separate classification accuracy from human-like visual understanding and generalization.
Can you explain with a practical example how models can fail to understand objects semantically and instead focus on superficial cues?
Asked once · last asked Aug 12 · Concept · Mid-level
Show answer
A model can learn a visual shortcut instead of learning what defines an object. For example, a cow classifier trained mostly on cows in grassy fields may use grass as its main clue. It may then miss a cow on a beach or label an empty field as a cow.
- This is shortcut learning: the model uses an easy clue that happens to match training labels.
- High test accuracy can hide this problem if test images contain the same clues.
- Test with changed backgrounds and train with more varied examples to reduce reliance on background.
What the interviewer checks: The interviewer is checking whether you can explain shortcut learning and how to reveal it with practical tests.
Computer vision fundamentals interview questions
Fundamentals questions cover images, color, filters, camera geometry, OpenCV, CNNs, transformers, and the limits of machine perception. 27 questions.
Can you explain the concept of the receptive field and why it is important for tasks like object detection or segmentation?
Asked once · last asked Aug 27 · Concept · Mid-level
Why can Python appear slow for OpenCV tasks, and where is the image-processing work actually executed?
Asked once · last asked Sep 11 · Concept · Senior
If you had to calculate the focal length of a camera, how would you do that?
Asked once · last asked Sep 23 · Scenario · Mid-level
What are the differences between using CCTV cameras and dedicated cameras for a computer vision application?
Asked once · last asked Oct 15 · Concept · Mid-level
What is OpenCV used for?
Asked once · last asked Oct 23 · Concept · Junior
Other ways it was asked
- “In what kinds of applications would you use OpenCV?”
What is a CNN used for?
Asked once · last asked May 12 · Concept · Junior
Can you give me examples of CNN models?
Asked once · last asked May 12 · Concept · Junior
What is bicubic interpolation?
Asked once · last asked May 19 · Concept · Junior
What is an image?
Asked once · last asked Jun 10 · Concept · Junior
How does a computer represent an image numerically?
Asked once · last asked Jun 10 · Concept · Junior
Other ways it was asked
- “How does a computer convert an image into a number?”
How is a color represented inside a computer?
Asked once · last asked Jun 10 · Concept · Junior
How does a computer learn what an eye, a face, or hair is?
Asked once · last asked Jun 10 · Concept · Junior
Show 5 more
Can you explain CNN?
Asked once · last asked Aug 2 · Concept · Junior
How do you tokenize image data for a computer vision model?
Asked once · last asked Aug 3 · Concept · Mid-level
What kind of neural network would you use for image classification?
Asked once · last asked Sep 3 · Concept · Mid-level
What is motion blur and what is median blur?
Asked once · last asked Sep 23 · Concept · Junior
What is a ViT (Vision Transformer)?
Asked once · last asked Oct 14 · Concept · Junior
Computer vision datasets and annotation interview questions
Questions cover collecting, storing, preparing, and labeling data, including annotation rules, consistency across video frames, and quality checks. 36 questions.
You’re reviewing bounding boxes for a moving object across frames and notice the annotations lag by three frames. How would you correct it while preserving temporal consistency?
Asked once · last asked Oct 3 · Scenario · Senior
Other ways it was asked
- “When bounding-box annotation lag varies from one frame early during slow motion to four frames late during fast motion, how would you correct it without causing box popping between frames?”
What is data annotation, and what types of image annotation are you familiar with?
Asked once · last asked Apr 22 · Concept · Junior
What is polygon annotation, and how does it differ from bounding-box annotation?
Asked once · last asked Apr 22 · Concept · Junior
What are Labelbox, Scale AI, and CVAT used for in image annotation?
Asked once · last asked Apr 22 · Tooling · Junior
What is the difference between frame-level and clip-level video annotation?
Asked once · last asked Jul 16 · Concept · Junior
What is the difference between bounding boxes, segmentation, and keypoint annotations?
Asked once · last asked Jul 16 · Concept · Mid-level
How would you handle a difficult video-annotation edge case such as occlusion or motion blur?
Asked once · last asked Jul 16 · Scenario · Mid-level
If a video has two people overlapping for a few seconds, one is partly blocked, how would you label it?
Asked once · last asked Jul 16 · Scenario · Mid-level
How would you keep a person's track consistent across frames when two people overlap and one is partly blocked?
Asked once · last asked Jul 16 · Scenario · Mid-level
If a pedestrian is visible for only three frames before exiting the scene, how would you keep the labels consistent across those frames?
Asked once · last asked Jul 16 · Scenario · Mid-level
When training YOLO to detect screws and a marker, would you label both object classes or only the screws?
Asked once · last asked Jul 21 · Concept · Mid-level
When you're annotating video frames to ensure consistent landmarks, how do you handle discrepancies or inconsistencies you might come across, and what steps do you take to resolve them?
Asked once · last asked Aug 23 · Scenario · Mid-level
Show 22 more
In a labeled video dataset, how would you detect incorrect systematic annotator drift over time, such as walking gradually being labeled as running? What evidence would you require before changing past labels?
Asked once · last asked Aug 24 · Scenario · Senior
A person starts running, passes behind a pole for 10 frames, then reappears still running. One continuous event or split, and what timestamp rule would you apply?
Asked once · last asked Aug 24 · Scenario · Mid-level
If your annotation tool can't represent hidden or occluded frames within an event, what exact rule would you use to decide when to split versus keep one event?
Asked once · last asked Aug 24 · Scenario · Senior
How would you preprocess, clean, and augment a large, noisy, or poorly labeled image dataset to improve model quality and robustness?
Asked once · last asked Aug 27 · Behavioral · Mid-level
Other ways it was asked
- “Tell us about a time you had to work with a large, noisy, or poorly labeled image dataset. What steps did you take to preprocess, clean, and augment the data to ensure the quality and robustness of your model?”
How would you process image data into ML-ready features, from raw images through the preprocessing and feature-extraction pipeline?
Asked once · last asked Aug 27 · Concept · Mid-level
Other ways it was asked
- “How did you process the image data—can you walk me through the pipeline of converting image data into ML-ready features?”
If a video is ambiguous because of partial occlusion or unclear guidelines, how do you decide whether to label it, mark it uncertain, or leave it unlabeled, and how do you justify that choice?
Asked once · last asked Sep 9 · Scenario · Senior
In a fast-moving scene with multiple overlapping actions, how would you choose boundaries so start and end times stay consistent across instances without chopping too finely or merging distinct actions?
Asked once · last asked Sep 9 · Scenario · Senior
How would you collect data for an automated driving test evaluation system?
Asked once · last asked Sep 10 · Scenario · Mid-level
In video data annotation, what is the difference between frame-level and track-level labeling, and when would you choose each?
Asked once · last asked Sep 11 · Concept · Mid-level
When you audit a tracking dataset, which quantitative metrics do you rely on to detect ID switches and temporal inconsistency, and what thresholds typically trigger rework?
Asked once · last asked Sep 11 · Scenario · Senior
How do you determine whether a spike in ID switches after a guideline update comes from guideline ambiguity, tooling, or annotator behavior?
Asked once · last asked Sep 11 · Scenario · Senior
In an image annotation project, how would you define a rule for borderline bounding boxes, specifically whether to include occluded or truncated parts, and how would you make it consistent across annotators?
Asked once · last asked Sep 14 · Scenario · Senior
When would you use image augmentation to improve a computer vision model?
Asked once · last asked Sep 22 · Concept · Mid-level
How would you consistently label segmentation boundaries when blur or low contrast makes pixels uncertain?
Asked once · last asked Oct 3 · Scenario · Senior
Other ways it was asked
- “In a segmentation mask review, you find consistent halo spill onto the background along hair and motion blur. What correction rule would you apply to keep masks accurate yet consistent?”
- “When the boundary is ambiguous due to heavy blur or low contrast, what rule do you use to include versus exclude uncertain pixels, and how do you apply it consistently across frames?”
What would you annotate if you were sitting in the passenger seat of a driverless car? What are the five parameters you would observe from the passenger seat?
Asked once · last asked Oct 13 · Scenario · Junior
What is key-point labeling?
Asked once · last asked Apr 22 · Concept · Junior
What preprocessing steps would you apply to images before using them in a computer vision model?
Asked once · last asked Jul 16 · Concept · Mid-level
How would you choose a storage solution for image data in a computer vision project?
Asked once · last asked Aug 3 · Scenario · Mid-level
What tools can be used to annotate images?
Asked once · last asked Aug 20 · Tooling
How should a briefly occluded event be represented in a video annotation?
Asked once · last asked Aug 24 · Scenario · Mid-level
What is the MNIST dataset, and what is it commonly used for?
Asked once · last asked Sep 3 · Concept · Junior
What input data do YOLO models use for training?
Asked once · last asked Sep 29 · Concept · Junior
Computer vision model training and evaluation interview questions
Model training and evaluation questions cover training loops, fine-tuning, losses, regularization, dataset splits, metrics, tuning, leakage, and model failures. 13 questions.
What strategies would you use for hyperparameter tuning in object detection models?
Asked once · last asked May 22 · Scenario · Mid-level
Suppose you're building a simple cats-and-dogs classification problem and achieve 90% accuracy on your local pipeline, but accuracy drops to 65% when the model sees real production data. What might be causing this, and how would you address it?
Asked once · last asked May 30 · Scenario · Mid-level
Tell me about a training or fine-tuning failure you encountered with YOLOv8 and how you fixed it.
Asked once · last asked Aug 12 · Behavioral · Mid-level
How would you implement a PyTorch training loop for a YOLOv8 model, including loading labels, computing losses, and diagnosing incorrect training or validation results?
Asked once · last asked Aug 12 · Tooling · Mid-level
How would you prevent data leakage or frame dependence from making a confidence interval for a model's precision improvement too optimistic?
Asked once · last asked Aug 12 · Scenario · Senior
Can you explain the full fine-tuning pipeline for the YOLO car-detection model?
Asked once · last asked Sep 7 · Scenario · Mid-level
How would you train an object-detection model end to end, from data collection and annotation through evaluation?
Asked once · last asked Sep 11 · Scenario · Mid-level
How do you fine-tune a computer vision model?
Asked once · last asked Sep 23 · Concept · Mid-level
If you had to train a model for roadside object detection, how would you approach it to keep it lightweight and avoid false negatives and false positives?
Asked once · last asked Sep 23 · Design · Senior
How do you choose a train, validation, and test split for a computer vision dataset?
Asked once · last asked Sep 29 · Scenario · Mid-level
How do you choose a loss function for training a computer vision model?
Asked once · last asked Sep 29 · Concept · Mid-level
Can you explain L1 and L2 regularization mathematically, from a computer vision perspective?
Asked once · last asked Oct 9 · Concept · Senior
Show 1 more
How do you train an object-detection model?
Asked once · last asked Sep 11 · Scenario · Mid-level
Object detection, segmentation, and tracking interview questions
Detection, segmentation, and tracking questions cover detector choices, YOLO, bounding boxes, masks, overlap metrics, non-maximum suppression, motion analysis, and tracking algorithms. 32 questions.
How would you evaluate an object detector, including its precision and recall?
Asked once · last asked Sep 7 · Scenario · Mid-level
Other ways it was asked
- “Which metrics would you use to evaluate an object-detection model?”
- “How would you evaluate the performance of an object-detection model?”
- “How do precision and recall help evaluate an object-detection model?”
How would you determine the region of interest for seat-belt detection?
Asked once · last asked Sep 11 · Scenario · Mid-level
Other ways it was asked
- “How would you extract the region of interest for seat-belt detection?”
- “What approach could you use for seat-belt detection instead of relying on a fixed region of interest?”
How would you handle variation in screw locations across supplier images when building an image-segmentation model?
Asked once · last asked Jul 21 · Scenario · Senior
Other ways it was asked
- “Would you still use an image-segmentation model when screw locations vary across supplier images?”
What parameters would you consider when comparing MobileNet SSD and YOLO?
Asked once · last asked Sep 16 · Concept · Mid-level
Other ways it was asked
- “What parameters would you consider when choosing between MobileNet-SSD and YOLO?”
How can you distinguish motion caused by the camera or lens from motion caused by objects in a video?
Asked once · last asked Apr 7 · Concept · Mid-level
How would you choose an object detection model for a given task?
Asked once · last asked Jul 16 · Concept · Senior
Other ways it was asked
- “How would you choose an architecture for an object detection task?”
- “Which model would you choose for an object detection task, and why?”
What does an object-detection model return when you give it an image?
Asked once · last asked Jul 21 · Concept · Junior
How would you calculate the centroid of each object from binary masks of markers and screws?
Asked once · last asked Jul 21 · Concept · Mid-level
How would you define a region of interest for calculating the centroid of each object in a binary mask?
Asked once · last asked Jul 21 · Scenario · Mid-level
How would you use bounding boxes to determine whether a marker overlaps with or is sufficiently close to a screw?
Asked once · last asked Jul 21 · Scenario · Mid-level
What metric would you use to measure the overlap between two bounding boxes?
Asked once · last asked Jul 21 · Concept · Junior
How does YOLOv8 differ from YOLOv4 or YOLOv5 in its backbone, training pipeline, and detection head, and what is one concrete engineering consequence of those differences?
Asked once · last asked Aug 12 · Concept · Senior
Other ways it was asked
- “How did YOLOv8 differ from YOLOv4 or YOLOv5 in your hands-on work, especially in the backbone, training pipeline, and detection head? Give one concrete engineering consequence of those differences.”
Show 18 more
How can OpenCV image preprocessing or postprocessing preserve bounding-box coordinates, and what operation would you use?
Asked once · last asked Aug 12 · Tooling · Mid-level
How can adjusting a YOLO threshold or post-processing improve model performance, and how would you measure the before-and-after impact?
Asked once · last asked Aug 12 · Scenario · Senior
How do attention mechanisms specifically improve performance in video analytics and object detection?
Asked once · last asked Aug 27 · Concept · Senior
How does ByteTrack work?
Asked once · last asked Sep 7 · Concept · Mid-level
How would you build a person detection and tracking solution that triggers an alarm when someone stays in or crosses a monitored area?
Asked once · last asked Sep 7 · Design · Senior
How would you keep tracking identities correctly when two people cross and their bounding boxes overlap?
Asked once · last asked Sep 7 · Scenario · Senior
Why might you choose object detection rather than image classification for seat-belt detection?
Asked once · last asked Sep 11 · Concept · Mid-level
How can contour processing be used to detect vehicle movement in video?
Asked once · last asked Sep 11 · Scenario · Mid-level
How would you distinguish actual vehicle movement from noise in video?
Asked once · last asked Sep 11 · Scenario · Mid-level
How would you determine whether a vehicle has crossed a specified line or boundary in a video?
Asked once · last asked Sep 11 · Scenario · Mid-level
Why is object tracking useful when vehicles are already being detected?
Asked once · last asked Sep 11 · Concept · Mid-level
How would you prevent, detect, and correct ID switches when objects cross in crowded scenes?
Asked once · last asked Sep 11 · Scenario · Mid-level
Other ways it was asked
- “How would you handle ID switches in crowded scenes when objects cross?”
- “How would you minimize ID switches in crowded scenes and detect or correct them during QA?”
How would you distinguish pedestrians from riders in roadside footage if the model has only one person class?
Asked once · last asked Sep 23 · Scenario · Mid-level
How would you determine from video whether a person is standing or walking?
Asked once · last asked Sep 23 · Scenario · Mid-level
How do you choose which metrics are best suited for evaluating your segmentation models?
Asked once · last asked Oct 20 · Concept · Mid-level
How does object tracking work for cars?
Asked once · last asked Sep 7 · Concept · Mid-level
Other ways it was asked
- “How does object tracking work for cars, and what libraries can be used?”
What is IoU?
Asked once · last asked Sep 7 · Concept · Junior
What are some object detection algorithms besides YOLO?
Asked once · last asked Sep 16 · Concept · Junior
Face recognition and gaze detection interview questions
Face and gaze questions cover detection, FaceNet, embeddings, enrollment, matching thresholds, evaluation, facial landmarks, and drowsiness detection. 29 questions.
What criteria would you use to select FaceNet over other face recognition models?
Asked once · last asked Sep 22 · Scenario · Mid-level
Other ways it was asked
- “What benchmark metrics would you compare when selecting FaceNet over other face recognition models?”
What libraries or models can be used to build a facial-recognition attendance system, and when would you train a custom model?
Asked once · last asked Jul 16 · Concept · Mid-level
How would you integrate biometric matching with CNN models in a Swift iOS application?
Asked once · last asked Jul 1 · Scenario · Mid-level
How would you process image data to detect drowsiness?
Asked once · last asked Aug 27 · Scenario · Mid-level
How would you implement a face detection and recognition system?
Asked once · last asked Sep 7 · Scenario · Mid-level
What tools or libraries can be used to compare face embeddings?
Asked once · last asked Sep 7 · Tooling · Mid-level
Which distance do you use for face embedding comparison?
Asked once · last asked Sep 7 · Concept · Mid-level
How can a mirror-gazing detection feature determine whether to use eye gaze alone or combine eye and head movement?
Asked once · last asked Sep 11 · Scenario · Mid-level
Other ways it was asked
- “Can gaze detection rely on eye movement alone, or should it account for head movement separately?”
How would you set or adapt gaze-detection thresholds when a driver's head is continuously rotating?
Asked once · last asked Sep 11 · Scenario · Senior
How would you compensate for head movement when estimating gaze from pupil position?
Asked once · last asked Sep 11 · Scenario · Senior
Other ways it was asked
- “How would you reduce noise in eye-gaze detection when head movement affects the measured pupil position?”
- “How would you handle changes in pupil location caused by slight head movement?”
- “How can you reduce the effect of head rotation on gaze detection?”
Why would you use one-to-one verification rather than one-to-many identification in a face-recognition system?
Asked once · last asked Sep 11 · Concept · Mid-level
How would you select a decision threshold for a face-recognition system?
Asked once · last asked Sep 11 · Scenario · Senior
Show 15 more
How does the FaceNet model work?
Asked once · last asked Sep 22 · Concept · Mid-level
How can a FaceNet-based face recognition system identify a person whose images were not included in training data?
Asked once · last asked Sep 22 · Concept · Mid-level
In a FaceNet-based face recognition system, what reference embedding is a live face embedding compared against?
Asked once · last asked Sep 22 · Concept · Mid-level
Which metrics and deployment factors, such as latency, model size, false positives, false negatives, precision, and recall, matter when selecting a face recognition model?
Asked once · last asked Sep 22 · Concept · Senior
How would you assess whether a face recognition model's size is suitable for deployment?
Asked once · last asked Sep 22 · Scenario · Mid-level
How would you set an accuracy target for a face recognition system?
Asked once · last asked Sep 22 · Scenario · Mid-level
How would you determine whether a dataset is large enough to benchmark a face recognition model reliably?
Asked once · last asked Sep 22 · Scenario · Mid-level
Which image preprocessing and augmentation techniques, such as grayscale conversion, rotation, skewing, or flipping, would you use for face recognition?
Asked once · last asked Sep 22 · Concept · Mid-level
How should a face recognition system handle matching failures caused by changes in facial expression or camera angle?
Asked once · last asked Sep 23 · Scenario · Mid-level
When would you train a face recognition model from scratch rather than use an existing model?
Asked once · last asked Sep 23 · Scenario · Mid-level
How would you choose and evaluate a face recognition threshold across different age groups or demographic groups?
Asked once · last asked Oct 11 · Scenario · Senior
What components are involved in designing and implementing an end-to-end mobile liveness detection library?
Asked once · last asked Oct 24 · Design · Senior
How many facial landmarks does a typical face-landmark model detect?
Asked once · last asked Jul 16 · Concept · Junior
Would a fixed pupil-movement threshold work reliably when a driver's head is rotating?
Asked once · last asked Sep 11 · Concept · Mid-level
What is the difference between face detection and face recognition?
Asked once · last asked Sep 23 · Concept · Junior
OCR and vision-language model interview questions
Topics include text detection, document extraction, OCR validation, screen analysis, hiding private information, vision-language models, and retrieval across data types. 26 questions.
How would you approach screen analytics to determine what an agent is doing and how much time they spend in applications such as Outlook or Facebook?
Asked once · last asked Oct 7 · Design · Senior
Other ways it was asked
- “How would you use screen analytics to measure an agent's activity, time spent in applications, and metrics such as emails sent?”
How do you define and measure OCR integrity for an address field? What specific checks or metrics would you use to decide a batch is safe for downstream annotation?
Asked once · last asked Jul 16 · Scenario · Senior
When your OCR output is semantically valid but differs from ground truth, for example, normalized abbreviations or reordered components, how do you decide whether to auto-accept, auto-correct, or route to manual review?
Asked once · last asked Jul 16 · Scenario · Senior
How would you validate that an OCR engine update didn't regress accuracy across document types? What would your evaluation dataset and acceptance criteria look like?
Asked once · last asked Jul 16 · Scenario · Senior
How would you extract and process images embedded in PDF documents for an LLM application?
Asked once · last asked Jul 16 · Scenario · Mid-level
How can a multimodal RAG system represent text and images in a common format so both can be searched?
Asked once · last asked Jul 16 · Scenario · Senior
What is involved in developing a custom AI model for CAPTCHA recognition, and what outcomes would you measure?
Asked once · last asked Jul 16 · Scenario · Senior
If a PDF contains charts like pie charts, line charts, or heat maps, would OCR still work, and how would you handle that case?
Asked once · last asked Jul 17 · Scenario · Mid-level
What model or technique would you use to detect text in an image?
Asked once · last asked Jul 21 · Concept · Mid-level
What is the difference between the Intelligent Form Extractor and the Machine Learning Extractor?
Asked once · last asked Sep 24 · Concept · Mid-level
What is the role of anchors in the Form Extractor and Intelligent Form Extractor?
Asked once · last asked Sep 24 · Concept · Mid-level
What preprocessing steps are useful before extracting attributes from text and images?
Asked once · last asked Oct 6 · Scenario · Mid-level
Show 12 more
When evaluating text and image attribute extraction, should you use exact-match or semantic-match metrics?
Asked once · last asked Oct 6 · Concept · Mid-level
How would you redact sensitive information from screen recordings while preserving critical text detail?
Asked once · last asked Oct 7 · Scenario · Senior
How would you detect and locate credit card numbers in a complex CRM screenshot without losing text detail during image resizing?
Asked once · last asked Oct 7 · Scenario · Senior
Other ways it was asked
- “How would you detect and locate credit card numbers in a complex CRM screen while preserving text detail?”
How would you decide whether a YOLO model is appropriate for a screen analytics task?
Asked once · last asked Oct 7 · Concept · Senior
How can OCR and named entity recognition be used in an AI-powered document privacy automation system?
Asked once · last asked Oct 12 · Concept · Mid-level
How would you use a vision-language model to interpret complex engineering drawings?
Asked once · last asked Oct 14 · Scenario · Senior
How would you validate a vision-language model's interpretation of engineering drawings, and when would you use a manually curated gold set?
Asked once · last asked Oct 14 · Scenario · Senior
What is optical character recognition (OCR), and what is it used for?
Asked once · last asked May 14 · Concept · Junior
How do OCR engines work?
Asked once · last asked Sep 7 · Concept · Junior
How does image segmentation relate to optical character recognition (OCR)?
Asked once · last asked Oct 13 · Concept · Mid-level
What is a VLM?
Asked once · last asked Oct 14 · Concept · Junior
What is OCR?
Asked once · last asked Oct 23 · Concept · Junior
Real-time computer vision deployment interview questions
Deployment questions focus on camera streams, measuring inference performance, latency, throughput, optimization, device versus cloud choices, and video processing at scale. 17 questions.
What is your thought process for selecting a computer vision algorithm to process 20 real-time camera streams at 30 FPS on an edge device?
Asked once · last asked Sep 16 · Scenario · Senior
Other ways it was asked
- “What is your thought process for selecting an algorithm for real-time object detection and tracking across 20 camera streams at 30 FPS on an edge device?”
- “How would you select an algorithm for real-time inference across 20 camera streams at 30 FPS on an edge device?”
How would you extract frames from an RTSP stream for face recognition?
Asked once · last asked Sep 22 · Tooling · Mid-level
How would you ensure the AI models process video frames quickly enough to meet the two-second alert requirement, especially as the number of streams increases?
Asked once · last asked Mar 27 · Design · Senior
When optimizing a convolutional neural network, which layers would you target, and why?
Asked once · last asked Jul 16 · Concept · Mid-level
For a real-time video detector, how would you balance accuracy against latency, and what would you profile or test to catch data-integrity problems and accuracy regressions after an optimization?
Asked once · last asked Aug 12 · Scenario · Senior
Consider this AI-generated claim: “Increasing batch size always lowers real-time latency, and YOLOv8’s objectness loss should be tuned separately from classification loss.” What is technically wrong or incomplete, and how would you write concise feedback to the author?
Asked once · last asked Aug 12 · Scenario · Senior
If a computer vision system processes 30 frames per second, does that mean its per-frame latency is about 33 milliseconds?
Asked once · last asked Sep 11 · Concept · Mid-level
If a camera produces 30 FPS but your model processes only 10 FPS, what could you do to handle the throughput gap?
Asked once · last asked Sep 11 · Scenario · Senior
How would you benchmark single-frame inference time for YOLOv5 or YOLOv8?
Asked once · last asked Sep 16 · Scenario · Mid-level
If YOLO takes 110 milliseconds to process one frame, how many frames per second can it process?
Asked once · last asked Sep 16 · Junior
How would you decide whether to use a cloud-based solution or a GPU-based edge solution for computer vision inference?
Asked once · last asked Sep 16 · Scenario · Senior
What does end-to-end deployment entail for a computer vision model that runs only on a workstation?
Asked once · last asked Sep 22 · Scenario · Mid-level
Show 4 more
Why use onboard machine vision in an unmanned system instead of relying on a server?
Asked once · last asked Sep 24 · Scenario · Mid-level
How would you implement platform channels, ML model integration, and camera streaming in a Flutter app?
Asked once · last asked Oct 24 · Design · Senior
How would you handle iOS or Android platform constraints when implementing real-time camera processing with machine learning?
Asked once · last asked Oct 24 · Behavioral · Senior
Other ways it was asked
- “What was the hardest platform constraint you hit on iOS or Android while running real-time camera plus ML, and how did it shape your implementation?”
Where would you deploy a computer-vision inspection solution, and what deployment environment would you choose?
Asked once · last asked Jul 21 · Design · Mid-level
Medical imaging and satellite computer vision interview questions
Medical, satellite, astronomical, and thermal imaging questions cover reconstruction, data access, resolution challenges, and validation for each field. 26 questions.
What are the dominant sources of false positives in lunar impact-flash detection?
Asked once · last asked Apr 15 · Scenario · Mid-level
Other ways it was asked
- “How can a model distinguish false positives in lunar impact-flash detection?”
Could a lunar impact-flash detection pipeline be generalized to detect space-based events such as glints or debris?
Asked once · last asked Apr 15 · Scenario · Senior
Other ways it was asked
- “What changes would you make to adapt a lunar impact-flash detection pipeline to detect space-based events?”
How would you extract rooftop solar imagery and obtain satellite images of rooftop solar panels?
Asked once · last asked Feb 25 · Case · Mid-level
What models and datasets can be used to classify medical images, and how can a model distinguish between different conditions?
Asked once · last asked Apr 11 · Concept · Mid-level
For detecting space-based events such as glints or debris, how would you compare CNNs with LSTMs or transformers?
Asked once · last asked Apr 15 · Concept · Senior
Why might you choose deep learning over traditional computer-vision techniques for lunar impact-flash detection?
Asked once · last asked Apr 15 · Concept · Mid-level
For lunar impact-flash detection, what models might you consider, and how would you choose among them?
Asked once · last asked Apr 15 · Scenario · Mid-level
How would you address spatial resolution challenges and class imbalance when working with satellite imagery?
Asked once · last asked May 22 · Scenario · Mid-level
How would you verify that a marker is positioned over a screw and that the temperature difference does not exceed 5 degrees?
Asked once · last asked Jul 21 · Scenario · Mid-level
What type of model would you use to detect screws and markers in a thermal image and draw bounding boxes around them?
Asked once · last asked Jul 21 · Concept · Mid-level
Once you have a binary mask, how would you validate an object's position and check its temperature?
Asked once · last asked Jul 21 · Scenario · Mid-level
How would you extract and validate temperature information from a thermal image?
Asked once · last asked Jul 21 · Scenario · Mid-level
Show 14 more
How do you determine whether synthetic images accurately represent real oncology findings?
Asked once · last asked Jul 27 · Scenario · Senior
How would you perform feature engineering for a skin cancer detection application?
Asked once · last asked Jul 27 · Scenario · Mid-level
How would you identify suitable data sources for a satellite imagery project?
Asked once · last asked Aug 3 · Scenario · Mid-level
Which libraries and techniques can you use to retrieve satellite imagery?
Asked once · last asked Aug 3 · Tooling · Mid-level
For ship detection using Sentinel-1 data, which polarization would you prefer—VV, VH, HH, or HV—and why?
Asked once · last asked Sep 2 · Case · Mid-level
What are the biggest challenges in X-ray and CT image reconstruction?
Asked once · last asked Oct 5 · Concept · Senior
How would you use Python to process and analyze satellite imagery?
Asked once · last asked Oct 9 · Tooling · Mid-level
What types of models can be used to analyze satellite imagery?
Asked once · last asked Oct 9 · Concept · Mid-level
What challenges can arise when preparing the LUNA16 dataset for lung nodule segmentation, and how would you address them?
Asked once · last asked Oct 19 · Scenario · Mid-level
When you’re working with clinical imaging data, such as lung CT scans, how do you ensure the data is well prepared for model training and validation?
Asked once · last asked Oct 20 · Scenario · Mid-level
What APIs or other data sources can provide satellite imagery?
Asked once · last asked Jul 16 · Concept · Mid-level
What is the intended use case for a skin cancer detection application?
Asked once · last asked Jul 27 · Case · Mid-level
How would you parse satellite image data for a computer vision project?
Asked once · last asked Aug 3 · Scenario · Mid-level
What are common use cases for computer vision projects involving satellite imagery?
Asked once · last asked Aug 3 · Case · Mid-level
Computer vision system design interview questions
System design questions cover complete vision applications in manufacturing, retail, driving, surveillance, agriculture, and robotics, plus point clouds, project trade-offs, and outcomes. 20 questions.
How would you design a fabric quality-control system that detects stains, stitching issues, holes, and color inconsistencies, and prioritizes items for human review?
Asked once · last asked Oct 21 · Design · Mid-level
How would you choose a model for detecting defects in fabric?
Asked once · last asked Oct 21 · Scenario · Mid-level
Other ways it was asked
- “How would you validate a model that detects defects in fabric?”
- “How would you collect and prepare data to train a model for fabric defect detection?”
- “What metrics would you use to evaluate a fabric defect-detection model?”
How would you design a real-time anomaly detection platform that processes video from hundreds of cameras, identifies suspicious activities using AI models, and alerts human reviewers in under two seconds, addressing architecture, major components, scalability, low latency, and reliable alert delivery?
Asked once · last asked Mar 27 · Design · Senior
Other ways it was asked
- “Design a real-time anomaly detection platform that processes video streams from hundreds of cameras, identifies suspicious activities using AI models, and alerts human reviewers in under two seconds. Explain the system architecture, its major components, and how you would ensure scalability, low latency, and reliable alert delivery.”
How would you integrate sensors and imaging modules in a quadcopter and process data to analyze crop-health patterns in real time?
Asked once · last asked May 20 · Design · Senior
How would you design an end-to-end computer vision pipeline?
Asked once · last asked Aug 3 · Design · Senior
Can you describe a complex computer vision project you led or played a significant role in, the biggest technical challenge, how you collaborated with cross-functional teams, and the final outcome?
Asked once · last asked Aug 27 · Behavioral · Mid-level
Imagine you are tasked with developing a system to detect and classify defects on a manufacturing assembly line using a live video feed. Can you walk me through your end-to-end approach from data and annotation strategy to model selection, training, and deployment?
Asked once · last asked Aug 27 · Design · Senior
How do you decide between YOLO, an NCNN-based model, or a segmentation model from stakeholder requirements?
Asked once · last asked Sep 7 · Scenario · Senior
How can a computer vision system determine whether a person has touched the ground rather than relying only on knee angle?
Asked once · last asked Sep 10 · Concept · Mid-level
How would you design an end-to-end system to stabilize an automated driving test?
Asked once · last asked Sep 11 · Design · Senior
How would you implement a computer vision system for face recognition and detecting helmets and seat belts?
Asked once · last asked Sep 22 · Scenario · Mid-level
How would you gather and use feedback from operators and point-cloud data users when evaluating a classification model?
Asked once · last asked Sep 23 · Scenario · Mid-level
Show 7 more
Imagine you are part of a team building a driverless car. What are the five basic parameters you would provide as data input so it can understand its surroundings and drive safely?
Asked once · last asked Oct 13 · Scenario · Mid-level
How would you design an overall system for point clouds so a user can change shapes through natural language?
Asked once · last asked Oct 14 · Design · Senior
Given product images from aspirational fashion brands and a retailer's own catalog, how would you build a solution to identify emerging trends those brands offer that the retailer does not?
Asked once · last asked Oct 15 · Design · Mid-level
Other ways it was asked
- “How would you compare product images from other brands with a retailer's catalog to identify trends missing from its assortment?”
How do you perform defect detection and process the data in a computer vision system?
Asked once · last asked Oct 15 · Concept · Mid-level
How can LiDAR data support business use cases?
Asked once · last asked Jul 16 · Case · Mid-level
What input data would a model need to detect whether a person is wearing a seat belt?
Asked once · last asked Sep 11 · Concept · Junior
Can you explain the architecture of a fabric quality-control system and the reasoning behind your design?
Asked once · last asked Oct 21 · Design · Mid-level
Related interview questions
- Machine Learning Engineer interview questions
- AI Engineer interview questions
- Data Scientist interview questions
- AI Trainer interview questions
- Software Engineer interview questions
- Machine Learning interview questions
- Deep Learning interview questions
- Generative AI and LLM interview questions
- RAG interview questions
- Agentic AI interview questions
- MLOps interview questions
How this list is made
Every question here was asked by a real interviewer in a real job interview that a candidate ran MeetAssist in. After the call, the interviewer's questions are pulled from the transcript. Practice sessions, tests, exams and anything that was not a job interview are dropped.
Questions are rewritten so they make sense on their own: names of people, employers and clients are removed, and speech-to-text mistakes are fixed. Small talk, logistics and questions about one candidate's own résumé are left out. Rephrasings of the same question are merged, and the count is the number of different interviews it came up in.
The interviews come from candidates who use MeetAssist, so they lean towards software, data and AI roles: the shares describe these interviews, not every Computer Vision interview. Answers to the most asked questions are written and checked by an editor; a "common mistake" is shown only when reviews of real answers in these interviews show it, with the count.
75 interviews, 55 candidates, Feb 2026 to Oct 2026. Updated every month with new interviews.
Or add the Chrome extension