Top 30 Intermediate Computer Vision Interview Questions and Answers
30 intermediate multiple-choice Computer Vision interview questions, past the definitions: how the pieces fit together, what breaks in practice, and the trade-off behind a choice. They come from 30 bites in the Computer Vision library, the middle slice of the 135 Computer Vision interview questions in the library. 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.
Image/video models, diffusion, OCR, multimodal
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.
Question 1 of 30
How are the intrinsic parameters fx and fy in K derived from physical camera properties?
Show the answer
Answer: a · They equal the focal length divided by pixel width and height respectively
fx and fy convert the physical focal length into pixel units by dividing by pixel width and height, making C correct. A is a common misconception because K stores focal length in pixel units, not millimeters, and D is wrong since distortion is modeled outside the idealized pinhole matrix.
Read the full bite: Explain the pinhole camera model and intrinsic matrix K
Question 2 of 30
During checkerboard calibration, why must you capture many images of the board at different orientations rather than a single frontal shot?
Show the answer
Answer: b · Varied views provide the geometric diversity needed to solve stably for intrinsics and distortion
Multiple poses constrain the parameter estimation enough to recover intrinsics and distortion coefficients reliably. One frontal view is degenerate and underdetermines the solution; it has nothing to do with resolution.
Question 3 of 30
Why is demosaicing necessary after a Bayer sensor captures an image?
Show the answer
Answer: a · Because each photosite records only one color channel, leaving missing values to estimate.
Demosaicing is required because every photosite measures only a single color channel, so the missing two channels must be interpolated from neighbors. Option D represents the common misconception that Bayer pixels already contain complete RGB data.
Read the full bite: How does a Bayer filter capture color and what is demosaicing?
Question 4 of 30
Switching from RGB to YCbCr does not reduce uncompressed frame size, yet 4:2:0 YCbCr cuts bandwidth roughly in half. What best explains where the savings come from?
Show the answer
Answer: d · Separating luma from chroma allows chroma planes to be stored at lower spatial resolution because human eyes have lower color spatial acuity.
The YCbCr transform is lossless and does not reduce uncompressed size; savings come from chroma subsampling, which exploits the human visual system's lower spatial resolution for color versus brightness. Distractor B is wrong because the transform does not inherently use fewer bits per pixel—it merely enables efficient subsampling and quantization.
Read the full bite: Compare YCbCr and RGB. Why chroma subsampling for compression?
Question 5 of 30
Why does a median filter remove salt-and-pepper specks more cleanly than a Gaussian blur of similar size?
Show the answer
Answer: c · Extreme outlier pixels sort to the ends of the window and are never chosen as the median
Median selection ignores extreme values, so corrupted black or white pixels are discarded while edges stay sharp. Gaussian blur averages those outliers into the result, smearing the noise and blurring edges.
Question 6 of 30
Which statement best explains why the Sobel Gx kernel has its specific 3x3 weight pattern?
Show the answer
Answer: b · It factors into a horizontal central-difference filter and an orthogonal vertical smoothing filter.
The Sobel Gx kernel is separable into a horizontal central-difference row and a vertical smoothing column, which reduces noise while estimating the partial derivative. Distractor A is wrong because, although the smoothing weights loosely approximate a Gaussian, the kernel is separable and explicitly not rotationally invariant.
Read the full bite: How does the Sobel operator approximate image gradients for edge detection?
Question 7 of 30
When applying a 3x3 convolution to the top border of a bright photo, which padding mode preserves spatial dimensions while avoiding dark vignettes and flat streaking?
Show the answer
Answer: c · Reflect-padding, because it mirrors edge pixels to assume continuity across the boundary
Reflect-padding mirrors edge pixels to maintain continuity across the boundary, avoiding both the dark vignettes caused by zero-padding and the flat streaking caused by replicate-padding. Replicate-padding is tempting because it avoids darkening, but it creates frozen-edge artifacts by repeating the same pixel value outward.
Read the full bite: Zero-padding vs reflect vs replicate padding and their visual artifacts
Question 8 of 30
In SIFT, which step is specifically responsible for rotation invariance of the descriptor?
Show the answer
Answer: c · Assigning each keypoint a dominant gradient orientation and describing it relative to that
Computing the descriptor relative to a keypoint's dominant orientation makes it invariant to rotation. Scale-space extrema give scale invariance, and normalization mainly addresses illumination, not rotation.
Question 9 of 30
For a real-time feature tracker on a battery-constrained phone, why is ORB often chosen over SIFT?
Show the answer
Answer: b · ORB's binary descriptors and FAST keypoints are far cheaper to compute and match
ORB's FAST detector and short binary descriptors give real-time speed and a small footprint, ideal for mobile. SIFT is more accurate and robust but too costly; ORB trades some robustness for performance.
Question 10 of 30
Why does Lowe's ratio test reject a match whose nearest and second-nearest descriptor distances are nearly equal?
Show the answer
Answer: b · Near-equal distances signal an ambiguous match with no distinctive best candidate
A reliable correspondence should be clearly closer than any alternative; when the two best are tied, the feature is not distinctive and the match is likely wrong. An absolute distance threshold would not capture this relative ambiguity.
Read the full bite: Descriptor matching and Lowe's ratio test
Question 11 of 30
What information about an image does the standard Bag of Visual Words representation deliberately discard?
Show the answer
Answer: c · The spatial positions and arrangement of the features
BoVW counts word frequencies into a histogram and ignores where features appear, just like a text bag-of-words ignores word order. Vocabulary size and classifier choice are separate design decisions, not discarded image content.
Question 12 of 30
What extra information does computing the essential matrix require that the fundamental matrix does not?
Show the answer
Answer: c · The camera intrinsic parameters to normalize the coordinates
The essential matrix works in calibrated normalized coordinates, so it needs the camera intrinsics, whereas the fundamental matrix is estimated from pixel correspondences alone. Absolute scale remains unknown even with the essential matrix.
Read the full bite: Fundamental matrix versus essential matrix
Question 13 of 30
In an incremental SfM pipeline, what is the primary role of bundle adjustment?
Show the answer
Answer: d · To jointly refine all camera poses and 3D points by minimizing reprojection error
Bundle adjustment is the nonlinear optimization that simultaneously adjusts poses and points to reduce reprojection error and control drift. Densification is a later multi-view stereo step, and matching and seed selection happen before it.
Read the full bite: Incremental Structure from Motion pipeline
Question 14 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
Question 15 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
Question 16 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
Question 17 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
Question 18 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.
Question 19 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
Question 20 of 30
Why does Mask R-CNN replace RoIPool with RoIAlign in its pipeline?
Show the answer
Answer: c · To avoid quantization that misaligns features with mask pixels
RoIAlign uses bilinear interpolation instead of rounding coordinates, preserving spatial alignment that pixel-accurate masks require. RoIPool's quantization is fine for boxes but introduces misalignment that visibly degrades masks.
Read the full bite: How does Mask R-CNN do instance segmentation?
Question 21 of 30
What is the most fundamental reason a deep segmentation network produces coarse mask boundaries?
Show the answer
Answer: c · Repeated downsampling in the encoder discards fine spatial detail
Aggressive spatial downsampling destroys high-frequency edge information that naive upsampling cannot restore. Optimizer minima, batch norm, and softmax temperature do not explain the systematic loss of spatial resolution.
Read the full bite: How to improve coarse segmentation boundaries?
Question 22 of 30
Why is the 'small motion' assumption necessary in classical optical flow derivations?
Show the answer
Answer: a · It justifies a first-order Taylor expansion that linearizes the constraint
Small displacements let you approximate the brightness constancy equation with a first-order Taylor expansion, making it linear and solvable. Color invariance is the separate brightness-constancy assumption, and gradients remain essential.
Read the full bite: Brightness constancy and small-motion assumptions
Question 23 of 30
In a constant-velocity Kalman filter for box tracking, why include velocity in the state but not in the measurement?
Show the answer
Answer: d · The detector reports only position and size; velocity is inferred from the state's dynamics
A detector directly observes the box position and size each frame, not its velocity, so velocity is a hidden state estimated from successive predictions and updates. The other options misstate how the filter or measurements work.
Question 24 of 30
What is a key practical disadvantage of the two-stream network compared to a 3D CNN for action recognition?
Show the answer
Answer: b · Its temporal stream needs expensive optical flow precomputed offline
The temporal stream consumes precomputed optical flow, which is costly to compute and store, unlike a 3D CNN that learns motion directly from RGB. Two-stream networks do capture motion and are not inherently larger than 3D CNNs.
Read the full bite: 3D CNNs vs two-stream action recognition
Question 25 of 30
Why is appearance-based Re-ID needed when an object is occluded for a long duration in MOT?
Show the answer
Answer: d · Motion prediction grows stale, so identity must be matched by learned appearance features
Over a long gap a Kalman or IoU prediction drifts and cannot reliably match the reappearing object, so a learned appearance embedding is used to recover the original id. The other options misdescribe what occlusion does to the system.
Read the full bite: Re-identification in multi-object tracking
Question 26 of 30
Why does a standard ViT typically need more training data than a comparable CNN?
Show the answer
Answer: c · It lacks built-in locality and translation equivariance, so must learn them from data
A CNN bakes in locality and translation equivariance, strong image priors that reduce data needs, while a plain ViT must learn spatial structure from data, making it data hungry. Parameter count and batch size are not the underlying reason.
Question 27 of 30
How does Swin keep information flowing across window boundaries despite using local windowed attention?
Show the answer
Answer: c · It shifts the window partition by half a window in alternating blocks
Shifted windows in alternating blocks straddle the previous boundaries, connecting neighboring regions across layers while keeping cost linear. Reverting to global attention would reintroduce quadratic cost, defeating Swin's purpose.
Read the full bite: How Swin Transformer achieves linear attention
Question 28 of 30
What distinguishes cross-attention from self-attention in a VQA model?
Show the answer
Answer: b · Queries come from one modality while keys and values come from the other
Cross-attention draws queries from one modality and keys/values from another, letting, say, text tokens attend over image features. It still uses softmax and learned projections, and applies across both modalities, so the other options are false.
Read the full bite: Cross-attention for visual question answering
Question 29 of 30
Why do Vision Transformers require positional embeddings while CNNs do not?
Show the answer
Answer: a · Self-attention is permutation invariant, so patch order must be supplied explicitly
Self-attention treats tokens as an unordered set, so without positional embeddings a ViT cannot know where a patch was; a CNN encodes position implicitly via its fixed convolution grid. Positional embeddings do not address gradients or attention cost.
Question 30 of 30
Why does mode collapse occur during GAN training?
Show the answer
Answer: a · The generator can win locally by repeatedly producing outputs that fool the current discriminator
Mode collapse stems from adversarial dynamics: nothing forces the generator to cover all modes, so it collapses onto whatever fools the current discriminator. It is distinct from overfitting, which is memorizing real samples.
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.