Skip to content
tezvyn:

Top 30 Computer Vision Interview Questions and Answers

30 multiple-choice questions on Computer Vision, drawn from 30 bites out of the 30 tagged Computer Vision on Tezvyn. Answer them here or read straight down. Every question carries the correct option, why it is correct, and a link to the bite it came from.

30 questions. Pick an answer, or open “Show the answer” to read it.

Answers are graded in your browser. Nothing is saved, and no XP or streak is earned here. The app keeps score.

  1. Question 1 of 30

    For which reason are HSL and HSV models generally unsuitable for image analysis and computer vision tasks?

    Show the answer

    Answer: b · They lack perceptual uniformity, making color difference measurements unreliable.

    The card states that HSL and HSV are not perceptually uniform, meaning visual changes do not correspond consistently to numerical changes, which makes calculating color distance unreliable for algorithms. Other options are either incorrect or not the primary reason cited.

    Read the full bite: HSL and HSV: Intuitive Ways to Represent RGB Color

  2. Question 2 of 30

    What is the primary reason a 2D Gaussian blur can be computed in O(N²K) rather than O(N²K²) for an N×N image and K×K kernel?

    Show the answer

    Answer: d · The Gaussian kernel is a rank-one matrix expressible as the outer product of two 1D vectors

    The Gaussian kernel is separable because it equals the outer product of two 1D Gaussians (a rank-one matrix), so two O(N²K) 1D passes replace one O(N²K²) 2D convolution. Using FFT is a distinct optimization, and neither normalization nor circular symmetry implies that a kernel can be decomposed into 1D passes.

    Read the full bite: How does filter separability optimize Gaussian blur and its complexity?

  3. Question 3 of 30

    Why does Canny use gradient orientation during non-maximum suppression?

    Show the answer

    Answer: b · To determine which neighboring pixels to compare for ridge thinning

    Gradient orientation tells NMS which neighbors lie along the edge direction so it can thin multi-pixel ridges to single-pixel width. Option C describes hysteresis, which uses spatial connectivity rather than gradient direction to link edges.

    Read the full bite: Walk me through Canny edge detection and why it beats Sobel thresholding

  4. Question 4 of 30

    In incremental Structure from Motion, what makes PnP the right tool for adding a new camera after the initial pair?

    Show the answer

    Answer: c · It estimates the new camera's pose by matching its 2D features to existing 3D points

    PnP recovers pose from known 3D-to-2D correspondences, registering each new frame into the existing cloud. The essential-matrix and joint-refinement options describe two-view geometry and bundle adjustment respectively, not single-camera registration.

    Read the full bite: The PnP problem in Structure from Motion

  5. Question 5 of 30

    Why does stereo rectification make corresponding points fall on the same image row?

    Show the answer

    Answer: a · It warps both images so epipolar lines become horizontal and the epipoles move to infinity

    Rectification applies homographies that make the image planes coplanar and parallel to the baseline, sending epipoles to infinity so epipolar lines are horizontal rows. Distortion removal and cropping are separate steps that do not by themselves align epipolar lines.

    Read the full bite: Stereo rectification math and its artifacts

  6. Question 6 of 30

    Which property of a convolutional layer most directly explains why it has far fewer parameters than a dense layer over the same input?

    Show the answer

    Answer: c · The same kernel weights are reused at every spatial position

    Weight sharing means one small kernel is applied across all positions, so parameters scale with kernel size not input size. Padding and stride affect output dimensions, not parameter count, and full connectivity is exactly what conv layers avoid.

    Read the full bite: How a convolutional layer works

  7. Question 7 of 30

    What is a key advantage of a strided convolution over max pooling for downsampling?

    Show the answer

    Answer: d · It learns how to downsample instead of fixing a non-learnable rule

    Strided convolution learns its downsampling jointly with feature extraction, unlike the fixed max rule. It actually adds parameters rather than saving them, and neither method guarantees full translation invariance.

    Read the full bite: Max pooling versus strided convolution

  8. Question 8 of 30

    If every activation function were removed from a deep CNN, what would happen to its representational power?

    Show the answer

    Answer: d · It would collapse to the power of a single linear layer

    A composition of purely linear operations is itself one linear map, so any depth reduces to a single linear layer that can only model linear boundaries. Nonlinearity, not depth alone, is what grants expressive power.

    Read the full bite: Why CNNs need nonlinear activations like ReLU

  9. Question 9 of 30

    An engineer switches from BFMatcher to FlannBasedMatcher in an OpenCV ORB pipeline and preserves the normType and crossCheck arguments. What is the most likely outcome?

    Show the answer

    Answer: a · FLANN silently ignores the unsupported Brute-Force parameters and uses its own defaults, producing unexpected matches without warning.

    The card warns that engineers often assume FLANN shares Brute-Force's normType and crossCheck parameters, causing silent configuration errors rather than exceptions or automatic adaptation. Option C is tempting because invalid arguments usually raise errors, but the source emphasizes that this mismatch fails silently.

    Read the full bite: FLANN Matcher for Feature Correspondence

  10. Question 10 of 30

    By what mechanism does dropout reduce overfitting during CNN training?

    Show the answer

    Answer: c · It randomly zeros activations so neurons cannot co-adapt, mimicking an ensemble

    Dropout deactivates random units each pass, breaking co-adaptation and effectively averaging many subnetworks. Penalizing weights is weight decay, transforming images is augmentation, and halting on validation loss is early stopping.

    Read the full bite: Regularization techniques for an overfitting CNN

  11. Question 11 of 30

    What problem did ResNet's residual blocks primarily solve in very deep networks?

    Show the answer

    Answer: b · The degradation problem where adding layers raised even training error

    Residual shortcuts let extra layers default to identity, so depth no longer degrades training accuracy. Overfitting concerns test error, not training error, and ResNet's contribution is optimizability, not inference speed or class balance.

    Read the full bite: ResNet residual blocks and the degradation problem

  12. Question 12 of 30

    Which factor most strongly multiplies the receptive field of a deep CNN neuron rather than just adding to it?

    Show the answer

    Answer: c · Stride and pooling that downsample the spatial resolution

    Stride and pooling downsample so each later step spans many more input pixels, multiplying the receptive field. Channel count, batch size, and dropout do not change the spatial region a neuron can see.

    Read the full bite: Receptive fields in convolutional networks

  13. Question 13 of 30

    What does a 1x1 convolution actually compute at each spatial position?

    Show the answer

    Answer: c · A learned linear combination across all input channels

    A 1x1 conv mixes channels via a per-pixel learned weighted sum, changing depth while preserving spatial size. It touches no spatial neighbors, is not a max, and is far from an identity since its weights are learned.

    Read the full bite: Uses of the 1x1 convolution

  14. Question 14 of 30

    Why are skip connections added when converting a classification CNN into a segmentation network?

    Show the answer

    Answer: a · To restore fine spatial detail lost during encoder downsampling

    Downsampling builds semantics but discards precise localization, so skip connections inject high-resolution encoder features into the decoder for sharper boundaries. They are not primarily about parameters, receptive field, or regularization.

    Read the full bite: Adapting a classification CNN for segmentation

  15. Question 15 of 30

    What assumption makes depthwise separable convolution a reasonable replacement for a standard convolution?

    Show the answer

    Answer: d · That spatial and cross-channel correlations can be learned separately

    The factorization splits spatial filtering from channel mixing, assuming these correlations are roughly independent. It does not assume identical channels, few channels, or that spatial structure is irrelevant, indeed depthwise still filters spatially.

    Read the full bite: Depthwise separable convolution cost savings

  16. Question 16 of 30

    Where does a classification CNN's approximate translation invariance actually come from?

    Show the answer

    Answer: d · From pooling and global aggregation that discard spatial position before the prediction

    Convolution is equivariant, not invariant; invariance arises only when pooling and global aggregation collapse spatial location before the classifier. Activations and kernel size do not produce shift-invariance.

    Read the full bite: Translation equivariance versus invariance in CNNs

  17. Question 17 of 30

    What does object detection provide that image classification does not?

    Show the answer

    Answer: a · Localization of multiple objects via bounding boxes with class labels

    Detection adds localization and handles multiple objects, outputting boxes with labels and scores. A single whole-image label is classification, and pixel-level masks describe segmentation rather than detection.

    Read the full bite: Image classification versus object detection

  18. Question 18 of 30

    How is Intersection over Union computed for a predicted and ground-truth bounding box?

    Show the answer

    Answer: c · Intersection area divided by the union of both boxes' areas

    IoU divides the overlapping area by the combined area of both boxes, ranging zero to one. Dividing by only the predicted area, inverting the ratio, or using center distance all describe different, incorrect metrics.

    Read the full bite: Intersection over Union for detection

  19. Question 19 of 30

    What makes COCO's primary mAP metric stricter than a single-threshold detection score?

    Show the answer

    Answer: a · It averages Average Precision over multiple IoU thresholds, rewarding tight localization

    COCO averages AP across IoU thresholds from 0.5 to 0.95, so loose boxes are penalized. It does not threshold on confidence at 0.95, drop rare classes, or use pixel masks for the box-based metric.

    Read the full bite: Mean Average Precision in object detection

  20. Question 20 of 30

    How does focal loss prevent easy background examples from dominating a dense detector's training?

    Show the answer

    Answer: a · It multiplies cross-entropy by a factor that shrinks the loss of well-classified examples

    The (1 - p)^gamma modulating factor smoothly suppresses confident, easy examples so hard ones drive learning. It does not delete anchors, and fixed per-class weighting is the separate alpha term, not the difficulty-based focal mechanism.

    Read the full bite: Focal Loss and class imbalance in detectors

  21. Question 21 of 30

    When optimizing a detector for a compute-limited edge device, why is on-device latency profiling more reliable than FLOP counts?

    Show the answer

    Answer: d · Real latency reflects memory bandwidth, operator support, and hardware quirks that FLOPs ignore

    Actual speed depends on memory bandwidth, supported operators, and accelerator behavior that raw FLOP counts do not capture. FLOPs neither consistently overestimate speed nor become unmeasurable after quantization, and profiling measures speed, not accuracy.

    Read the full bite: Deploying real-time detection on edge devices

  22. Question 22 of 30

    Why is Smooth L1 commonly chosen over pure L2 loss for bounding box regression?

    Show the answer

    Answer: d · It is linear for large errors, keeping gradients bounded and robust to outliers

    Smooth L1 is quadratic near zero but linear for large errors, so outlier coordinates do not produce exploding gradients as L2 would. It is a regression loss, applies to positive anchors, and offers no exactness guarantee.

    Read the full bite: Detector head losses: regression versus classification

  23. Question 23 of 30

    What can instance segmentation do that semantic segmentation cannot?

    Show the answer

    Answer: c · Separate individual objects of the same class into distinct masks

    Instance segmentation distinguishes separate objects of the same class, which semantic segmentation merges into one labeled region. Per-pixel class labeling is what both share, and boxes or whole-image labels describe detection and classification.

    Read the full bite: Semantic versus instance segmentation

  24. Question 24 of 30

    What is the specific role of U-Net's skip connections between encoder and decoder?

    Show the answer

    Answer: d · They concatenate high-resolution encoder features into the decoder to restore spatial detail

    U-Net concatenates resolution-matched encoder features into the decoder, reinjecting fine spatial detail lost to downsampling. They are concatenations, not residual additions, and they supplement rather than replace upsampling.

    Read the full bite: U-Net architecture and its skip connections

  25. Question 25 of 30

    Why does the Lucas-Kanade method assume all pixels in a small window share the same motion?

    Show the answer

    Answer: a · To resolve the aperture problem by making the system overdetermined

    A single pixel gives one equation in two unknowns (the aperture problem); assuming a window of pixels moves identically yields an overdetermined system solvable by least squares. It does not concern grayscale conversion or removing gradients.

    Read the full bite: Sparse vs dense optical flow and Lucas-Kanade

  26. Question 26 of 30

    For which application is U-Net the most appropriate architecture?

    Show the answer

    Answer: c · Tracing precise cell boundaries in 512x512 microscopy images with only 200 training masks

    U-Net is built for pixel-accurate segmentation when training data is limited, exactly as described in option C. Option B is simple classification, B is bounding-box detection on high-resolution footage, and D combines massive data with only coarse localization needs, all scenarios the card identifies as poor fits for U-Net.

    Read the full bite: U-Net: Segmentation with Less Data

  27. Question 27 of 30

    In a CNN built for image classification, what is the primary role of pooling layers placed between convolutional blocks?

    Show the answer

    Answer: a · They reduce spatial dimensions and introduce translation invariance using fixed operations

    Pooling layers use fixed operations such as max or average pooling to shrink spatial dimensions and provide translation invariance, and they have no learnable parameters. Option B is a tempting distractor because many candidates mistakenly believe pooling layers learn adaptive weights, when in fact they perform static downsampling.

    Read the full bite: Walk me through a CNN's layers for image classification

  28. Question 28 of 30

    A road-scene model labels every pixel as the dominant background class. How will its pixel accuracy and mIoU compare?

    Show the answer

    Answer: c · High pixel accuracy but low mIoU

    Predicting only background scores well on pixel accuracy because background dominates, but every foreground class gets zero IoU, dragging mIoU down. The option claiming both are high ignores that minority classes are never predicted.

    Read the full bite: How is IoU computed and why prefer mIoU?

  29. Question 29 of 30

    In a 5-way 3-shot episode, how many labeled examples are in the support set before the model classifies the query images?

    Show the answer

    Answer: c · 15

    The support set holds N times K, which is five times three, equal to fifteen labeled examples. The query set size is separate and does not change the support count, so the last option conflates the two distinct sets.

    Read the full bite: What does N-way K-shot classification mean?

  30. Question 30 of 30

    Why do engineers choose an FPGA over a GPU for a real time vision pipeline with a fixed algorithm?

    Show the answer

    Answer: c · Because a fixed pipeline can be built as dedicated parallel hardware, giving deterministic low latency and low power instead of scheduling instructions on shared cores

    FPGAs win when the workload is fixed and latency and power predictability matter, since the algorithm becomes physical hardware instead of instructions competing for a shared scheduler. Cost and ease of development actually favor GPUs, which is why A is the only accurate tradeoff.

    Read the full bite: FPGA in Computer Vision

Could you explain these out loud?

That is what an interview actually tests. Tezvyn gives you questions like these with what the interviewer is really checking, the answer that lands, and the mistake that ends the conversation, in the four minutes before your next meeting.

The iPhone app is on the way

We are building it. Until it lands, nothing here is held back from you: every interview card, your saved cards, streaks and the job board all work in Safari, plus hundreds of free practice quizzes of thirty questions each. Sign in and it all carries over to the app the day it arrives.

Want it as an icon? Tap Share at the bottom of Safari, then Add to Home Screen. It opens full screen and the cards you have read stay available offline.

Get it on Google PlayiPhone app coming soon