Top 30 Advanced Computer Vision Interview Questions and Answers
30 advanced multiple-choice Computer Vision interview questions, the deep end: internals, failure modes, and the design calls a senior engineer is expected to defend. They come from 30 bites in the Computer Vision library, the hardest 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
Why does a fixed Euclidean distance threshold in RGB fail to produce consistent perceptual segmentation across light and dark image regions?
Show the answer
Answer: a · The same numerical RGB delta can correspond to a huge perceived shift in one region while being nearly invisible in another
RGB is linear with respect to light intensity but not human perception, so identical Euclidean deltas can look huge in one region and nearly invisible in another. HSV is a cylindrical transform of RGB and is not perceptually uniform, while CIELAB is designed for perceptual uniformity rather than linear physical intensity.
Read the full bite: Why is RGB Euclidean distance a poor measure of perceptual color difference?
Question 2 of 30
When photometric stereo assumes a Lambertian model but the surface is glossy, false geometry is reconstructed because the solver...
Show the answer
Answer: c · misinterprets view-dependent specular brightness as a tilted surface normal
Under a Lambertian assumption, the solver expects brightness to depend only on the light direction and surface normal, so view-dependent specular highlights are misread as changes in normal orientation. Option A describes a common conceptual error—treating the BRDF as a scalar albedo—but it does not explain the specific mechanism that creates phantom geometry in photometric stereo.
Read the full bite: Describe the BRDF, its advantage over Lambertian, and critical CV tasks
Question 3 of 30
When rotating an image, why is inverse mapping preferred over forward mapping of pixels?
Show the answer
Answer: b · It guarantees every output pixel gets exactly one interpolated value, avoiding holes and overlaps
Iterating over output pixels and sampling the source ensures full, single-valued coverage, while forward mapping leaves gaps and collisions. Inverse mapping still uses the inverse transform and interpolation to read fractional source coordinates.
Read the full bite: Image rotation: forward versus inverse mapping
Question 4 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?
Question 5 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
Question 6 of 30
When repurposing a pretrained CNN like VGG16 for image retrieval, why are deep-layer activations usually preferred over the final softmax output?
Show the answer
Answer: a · Deep activations encode rich semantic features, while softmax collapses the image to class probabilities
A late hidden layer yields a high-dimensional semantic descriptor ideal for similarity comparison, whereas the softmax discards detail by reducing the image to class scores. Earlier layers still carry useful low-level information, just less semantic content.
Question 7 of 30
Why is ORB typically preferred over SIFT for the front end of a real-time visual SLAM system on a mobile device?
Show the answer
Answer: d · ORB's cheap FAST keypoints and binary descriptors meet the per-frame latency and power budget
Real-time SLAM needs detection, description, and matching within milliseconds at low power, which ORB's FAST keypoints and Hamming-matched binary descriptors satisfy. SIFT is more robust but too slow; temporal continuity offsets ORB's weaker invariance.
Read the full bite: Feature choice for real-time mobile SLAM
Question 8 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
Question 9 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
Question 10 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
Question 11 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
Question 12 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
Question 13 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
Question 14 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
Question 15 of 30
Panoptic Quality is the product of which two components, and what does each capture?
Show the answer
Answer: c · Segmentation Quality (mean IoU of matches) and Recognition Quality (F1 over segments)
PQ = SQ x RQ, where SQ is the average IoU of matched true positives and RQ is an F1 over segments. Plain precision/recall or mIoU do not separate mask tightness from detection correctness the way PQ does.
Read the full bite: Explain panoptic segmentation and Panoptic Quality
Question 16 of 30
What is a primary challenge when using a plain ViT encoder for high-resolution semantic segmentation versus a CNN?
Show the answer
Answer: b · Self-attention is quadratic in patch count and ViT lacks a native feature pyramid
Attention cost grows quadratically with the number of patches and a plain ViT keeps a single resolution, lacking the multi-scale pyramid CNNs provide. ViTs do handle color and can be adapted for dense output, so those options are false.
Read the full bite: Adapting ViT for dense semantic segmentation
Question 17 of 30
Which property must a good self-supervised pretext task for video have to benefit action recognition?
Show the answer
Answer: c · Solving it should require understanding temporal dynamics and motion
A useful pretext task forces the model to reason about how content changes over time, yielding features that transfer to action recognition. Single-frame-solvable tasks teach no motion, and self-supervision by definition uses no manual labels.
Read the full bite: Self-supervised pretraining for video understanding
Question 18 of 30
What fundamentally distinguishes scene flow from optical flow?
Show the answer
Answer: c · Scene flow is a 3D motion field requiring depth, not a 2D image-plane displacement
Scene flow is the dense three-dimensional motion of points in space and needs depth from stereo, RGB-D, or LiDAR, whereas optical flow is the two-dimensional apparent motion in the image plane. It is not limited to grayscale, background, or whole objects.
Question 19 of 30
How does a NeRF represent and render a 3D scene?
Show the answer
Answer: d · An MLP maps 3D point plus view direction to color and density, composited along rays by volume rendering
NeRF encodes the scene implicitly in an MLP that outputs color and density per point and view, then integrates these along camera rays via differentiable volume rendering. It does not store an explicit mesh, voxel grid, or point cloud.
Read the full bite: Core principles of a Neural Radiance Field
Question 20 of 30
Why is a hybrid CNN-Transformer often preferred over a pure ViT for high-resolution medical segmentation?
Show the answer
Answer: c · The CNN gives data-efficient local features and crisp boundaries while attention adds global context
Medical datasets are small and need both fine boundaries and global context; a CNN encoder supplies data-efficient local detail while transformer layers add long-range context, with decoder skips restoring resolution. ViTs can do dense output but are data hungry, and transformers do capture global context.
Read the full bite: Pure ViT vs hybrid CNN-Transformer for medical segmentation
Question 21 of 30
In a text-conditioned diffusion U-Net, which tensors supply the keys and values of the cross-attention layers?
Show the answer
Answer: a · The encoded text token embeddings
Cross-attention injects the prompt by using encoded text tokens as keys and values while image features act as queries. The noisy image features supply the queries, not the keys and values.
Read the full bite: Attention in diffusion U-Nets for text conditioning
Question 22 of 30
How is classifier-free guidance enabled during training?
Show the answer
Answer: c · By randomly dropping the conditioning signal so one model learns conditional and unconditional denoising
Randomly nulling the condition lets a single network learn both conditional and unconditional denoising, which are combined at inference. Training a separate noisy-image classifier describes classifier guidance, the method it replaces.
Read the full bite: Classifier-free guidance in diffusion models
Question 23 of 30
What is a key practical advantage of DDIM sampling over standard DDPM sampling?
Show the answer
Answer: b · It allows deterministic sampling on far fewer steps without retraining
DDIM reuses the same trained model but samples deterministically over a sparse step subset, cutting inference cost greatly. It needs no separate model and does not change image resolution.
Question 24 of 30
What is a primary reason to favor a diffusion model over a StyleGAN for high-resolution face generation?
Show the answer
Answer: a · More stable training and stronger mode coverage at the cost of slower sampling
Diffusion training is stable and covers modes well, reducing collapse, but sampling is slow due to many steps. Fast single-pass inference and built-in disentangled style control are strengths of StyleGAN, not diffusion.
Read the full bite: Designing a high-resolution photorealistic face generator
Question 25 of 30
Which is a genuine limitation of the original NeRF formulation for robotics?
Show the answer
Answer: c · It is slow to train and render and assumes a static scene
Vanilla NeRF is per-scene, static, pose-hungry, and slow because of dense MLP ray queries, which is why it is not real-time. It does not generalize across scenes and does need accurate poses.
Read the full bite: NeRF limitations and advances for robotics
Question 26 of 30
Why is a single sparse terminal reward problematic for this long-horizon robot task?
Show the answer
Answer: d · Credit assignment over many steps becomes intractable, so the agent rarely discovers the success path
With only a final reward, the agent almost never stumbles onto success across a long horizon, so it gets little learning signal. Shaped subgoal rewards make credit assignment tractable; this is unrelated to the Markov property or action discreteness.
Read the full bite: Formulating a multi-step robot manipulation task
Question 27 of 30
In MAML, what is actually being learned by the outer loop optimization?
Show the answer
Answer: c · A parameter initialization from which a few gradient steps adapt to new tasks
The outer loop optimizes the shared initialization so that brief task-specific inner-loop adaptation generalizes. It is not a single all-task solver, and unlike metric methods it does not learn a distance function.
Read the full bite: How does MAML's inner and outer loop work?
Question 28 of 30
Why does a plain Siamese network trained only to match two views of an image tend to collapse?
Show the answer
Answer: d · Outputting a constant trivially minimizes the agreement loss
With only an agreement objective, a constant output makes the loss zero, the degenerate solution. BYOL and Barlow Twins add asymmetry or decorrelation to block it; augmentation strength alone does not cause collapse.
Read the full bite: How do BYOL and Barlow Twins avoid representation collapse?
Question 29 of 30
What extra information does a light field record that a normal camera image discards?
Show the answer
Answer: c · The direction of incoming rays, not just their accumulated intensity
A light field captures both position and direction of rays, enabling post-capture refocus by reintegrating them. A normal sensor sums all directions into one intensity, losing that directional data.
Read the full bite: How does a plenoptic camera enable post-capture refocus?
Question 30 of 30
Roughly how does the signal-to-noise ratio improve when averaging N well-aligned burst frames?
Show the answer
Answer: b · It improves by about the square root of N
Independent noise averages down by the square root of the number of frames, so SNR rises as the square root of N. Linear or quadratic scaling overstates the gain, and the noise clearly does decrease.
Read the full bite: Why merge a burst instead of one long low-light exposure?
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.