In this article
How to use this list
Read an answer, say it aloud in your own words, then open the linked page for the full explanation. The list sits in the deep learning learning path, and the detection metrics lean on counting and ratios, which the probability and statistics guide covers. For more general network questions, see the deep learning interview questions. Each entry gives a short answer, one line on why it is asked, and a link to a fuller source. This page does not say that any named company asks a given question.
Convolution basics
What is a convolutional neural network, and what does a convolution layer do?
A convolutional neural network (CNN) is a network built mainly from convolution layers. A convolution layer is a layer whose neurons each connect to a small local patch of the input and span its full depth. A receptive field is that small patch one neuron reads. Connectivity is local in width and height but full along depth (Stanford CS231n notes, accessed October 2026).
Why it is asked: it exposes the misconception that convolution means every neuron sees the whole image. Read the full explanation: Stanford CS231n convolutional networks notes.
How do you compute the output size of a convolution layer?
The spatial output size is (W − F + 2P)/S + 1, where W is the input size, F the filter size, S the stride and P the zero padding (CS231n notes, accessed October 2026). A 7 by 7 input with a 3 by 3 filter, stride 1 and no padding gives 5 by 5; stride 2 gives 3 by 3. A 227 input with an 11 filter and stride 4 gives 55.
Why it is asked: it checks that you can do shape arithmetic on a whiteboard. Read the full explanation: CS231n notes.
What does a pooling layer do?
A pooling layer is a layer that shrinks each feature map by summarizing small windows, most often by keeping the maximum of each window. A 2 by 2 max-pool at stride 2 keeps one of every four activations, so it discards 75% of them (CS231n notes, accessed October 2026).1 That cuts computation and memory and adds some tolerance to small shifts, but it loses fine position detail.
Why it is asked: it checks that you know what downsampling costs. Read the full explanation: CS231n notes.
How does an edge detector work?
An edge detector finds places where pixel brightness changes sharply. It slides a small gradient filter, such as the Sobel filter, across the image: one filter responds to horizontal change, another to vertical change, and the combined strength is large at an edge. This is a convolution with fixed weights (illustrative kernel rows such as −1, 0, 1), where a CNN learns its weights instead.
Why it is asked: it links classical image processing to what a first convolution layer learns. Read the full explanation: CS231n notes.
What padding keeps the output the same size as the input?
With stride 1, a padding of P = (F − 1)/2 keeps the output the same size as the input. This follows from the output formula above: W − F + 2P + 1 equals W exactly when P is (F − 1)/2, so a 3 by 3 filter needs one ring of zeros.1 Odd filter sizes make that padding a whole number, which is one reason 3 by 3 filters are common.
Why it is asked: it tests whether you can derive a rule from a formula instead of memorizing it. Read the full explanation: CS231n notes.
Why can a stride and filter setup be impossible?
A setup is impossible when the formula gives a non-integer, because the filter would not fit a whole number of times. A 10-wide input with a 3-wide filter, stride 2 and no padding gives 4.5, so that layer cannot be built.1 Fix it by changing the padding, the stride or the filter size. Run the formula before you claim a layer works.
Why it is asked: it shows whether you check shapes before building. Read the full explanation: CS231n notes.
How many parameters does a convolution layer have, and why does sharing matter?
In the CS231n example, a 227-wide input, 11 by 11 filters, stride 4 and 96 filters give a 55 by 55 by 96 output: 290,400 neurons with 364 inputs each (363 weights plus a bias). Without sharing that is 105,705,600 parameters (CS231n notes). Sharing one filter across positions leaves 96 filters of 364 numbers, or 34,944 parameters, which is 3,025 times fewer.1
Why it is asked: parameter sharing is the main reason convolution scales to images. Read the full explanation: CS231n notes.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →Deep architectures
Why are very deep networks hard to train, and what do residual connections do?
A residual connection adds a block's input to the block's output, so the block only has to learn the residual F(x) = H(x) − x rather than the full mapping H(x). He et al. (2015) start from the observation that very deep plain networks are harder to train, and this reformulation eases that (He et al., 2015). They evaluate residual nets up to 152 layers on ImageNet, and an ensemble reached 3.57% test error, winning the 2015 classification task.
Why it is asked: it corrects the belief that deeper is automatically better. Read the full explanation: residual connections and the ResNet paper.
What do activation functions do in a convolutional network?
An activation function is a non-linear function applied to a layer's outputs. Without it, stacked layers collapse into a single linear map, so extra depth would add no expressive power. Convolutional networks commonly use the rectified linear unit and its variants. The choice affects how well gradients flow during training.
Why it is asked: it checks that you know why non-linearity is needed, not just the names. Read the full explanation: activation functions.
What is batch normalization, and how does it differ from layer normalization?
Batch normalization rescales each channel using the mean and variance of the current mini-batch, then applies a learned scale and shift. Layer normalization computes those statistics across the features of each single example, so it does not depend on batch size. Convolutional networks often use the first and transformers the second. Very small batches make batch statistics noisy. At inference, batch normalization uses running averages of the batch statistics collected during training, so a prediction does not depend on the other images in the batch.
Why it is asked: it separates memorized definitions from knowing when each one breaks. Read the full explanation: batch norm vs layer norm.
What are vanishing and exploding gradients, and how do you fix them?
Gradients shrink toward zero (vanish) or grow without bound (explode) as they pass back through many layers, which stalls or destabilizes training. Remedies include gradient clipping for explosions, careful weight initialization, normalization layers and residual connections. Name the symptom first: a loss that turns into not-a-number points to exploding gradients.
Why it is asked: it tests whether you can diagnose a failing training run. Read the full explanation: exploding gradients.
Transformers in vision
Do vision models need convolutions?
No. Dosovitskiy et al. (2020) show that a pure transformer fed sequences of image patches can do very well at image classification, so a convolutional backbone is not a requirement (Dosovitskiy et al., 2020). Earlier work used attention alongside convolutional networks or in place of some of their components. CNNs are not obsolete; a careful answer adds that the strong results are associated with large-scale pre-training.
Why it is asked: it checks whether you can state a research claim at its real strength. Read the full explanation: the Vision Transformer paper.
What is the attention mechanism in computer vision?
Attention is a mechanism in which each element of a sequence computes a weighted combination of the other elements, with weights set by relevance. In a Vision Transformer the elements are image patches, so one layer lets a patch draw on distant patches, where a convolution sees only its local receptive field. The cost is that the number of patch pairs grows with the square of the patch count.
Why it is asked: it tests whether you can contrast global and local context. Read the full explanation: the Vision Transformer paper.
What is a mixture of experts in a vision model?
A mixture of experts replaces one large sub-network with several expert sub-networks and a small router that sends each input, or each image patch, to a few of them. Only the chosen experts run, so capacity grows without all of it running every time. Sparsely activated vision transformers of this kind are described by Riquelme et al. (2021). The trade-offs are balancing load across experts and the memory needed to hold every expert's weights.
Why it is asked: it probes whether you understand conditional computation and its costs. Read the full explanation: feed forward networks, whose layers the experts replace.
Detection and metrics
What is intersection over union?
Intersection over union (IoU) is the area where a predicted box and a ground-truth box overlap, divided by the area they cover together. It runs from 0 for no overlap to 1 for a perfect match. In an illustrative example, two 10 by 10 boxes shifted 2 units apart overlap by 80 and cover 120 together, so the IoU is 0.6667.1 A detection of the right class is a true positive when its IoU with a ground-truth box that is not yet matched is at least the threshold. Detections are matched highest score first, so a duplicate on an already matched object is a false positive. The same box can pass at 0.5 and fail at 0.75.
Text description of the diagram
- A detection of the right class is a true positive when its IoU with a ground-truth box that is not yet matched is at least the threshold.
- Detections are matched highest score first, so a duplicate on an already matched object is a false positive.
Why it is asked: nearly every detection metric is built on it. Read the full explanation: Common Objects in Context (COCO) detection evaluation.
What are AP and mAP, and how does COCO compute them?
Average precision (AP) is a single score that summarizes how precision falls as recall rises for one class. The COCO benchmark averages it over 10 IoU thresholds, 0.50 to 0.95 in steps of 0.05, and over 80 categories, and that number decides the challenge winner. COCO makes no distinction between AP and mean average precision (mAP) (COCO detection evaluation, accessed October 2026).
Why it is asked: candidates often confuse the COCO metric with the older single-threshold one. Read the full explanation: COCO detection evaluation.
Why does COCO average over IoU thresholds?
Averaging over IoU thresholds rewards detectors with better localization. It departs from the older practice of computing AP at the single IoU of 0.50 (COCO detection evaluation, accessed October 2026). A detector with loose boxes can look fine at 0.50 but loses credit at the stricter thresholds, so the average separates tight boxes from loose ones.
Why it is asked: it checks that you can explain a metric's design, not only its name. Read the full explanation: COCO detection evaluation.
What does non-maximum suppression do?
Non-maximum suppression (NMS) is a post-processing step that removes duplicate detections of the same object. A detector emits many overlapping boxes; NMS keeps the highest-scoring one, deletes the others whose IoU with it exceeds a chosen threshold, and repeats. This matters for scoring, because a second high-overlap box on an object that is already matched counts as a false positive, so NMS raises precision. A threshold set too low deletes real neighbours in crowded scenes; too high leaves duplicates.
Why it is asked: it tests whether you know the step between raw detector output and a clean result. Read the full explanation: You Only Look Once (YOLO) and Darknet.
What is the difference between one-stage and two-stage detectors?
A two-stage detector first proposes regions likely to hold objects, then classifies and refines each; a one-stage detector such as You Only Look Once (YOLO) predicts boxes and classes directly in one pass. One-stage designs tend to favor speed, and two-stage designs have often favored accuracy, though the gap depends on the model and the era. State the trade-off, then say which one your latency budget needs.
Why it is asked: it checks whether you can match an architecture to a constraint. Read the full explanation: YOLO and Darknet.
How does COCO group objects by size?
COCO reports results by object size: small objects have an area below 32 squared, medium objects from 32 squared to 96 squared, and large objects above 96 squared (COCO detection evaluation, accessed October 2026). Those limits are 1,024 and 9,216 pixels of area, and COCO measures area as the pixels in the segmentation mask, not the box area.1 Report AP per size to see where a detector fails, since small objects keep few pixels after downsampling.
Why it is asked: an overall score can hide a weakness on the objects your product cares about. Read the full explanation: COCO detection evaluation.
What is the difference between semantic and instance segmentation?
Semantic segmentation labels every pixel with a class but treats all objects of one class as a single region; instance segmentation also separates each individual object. Mask R-CNN does the latter by extending Faster R-CNN with a branch that predicts a mask for each detected instance (He et al., 2017). Two people side by side share one label in semantic output and get two masks in instance output.
Why it is asked: it checks that you can match the output type to the product need. Read the full explanation: the Mask R-CNN paper.
Practical training
How do you fine-tune a pretrained vision model on a small dataset?
Fine-tuning is a form of transfer learning: start from a network pretrained on a large dataset, replace the final layer for your classes, and train first with the pretrained layers frozen, then with a small learning rate across more layers. Use augmentation such as flips and crops, and watch validation loss for overfitting. For class imbalance, reweight the loss or resample the rarer classes.
Why it is asked: it shows whether you can ship a model without a huge dataset. Read the full explanation: deep learning interview questions.
Follow-up questions
What are the common computer vision engineer interview questions?
The twenty-two above cover the recurring ground: shape arithmetic, parameter counts, residual networks, transformers in vision, IoU and average precision. Interviewers often ask for a number you can derive, so practise the formulas aloud.
What are senior computer vision engineer interview questions?
Senior questions usually go deeper on the same topics: why a metric was chosen, where a detector fails by object size, and what a latency budget forces you to trade. A strong answer shows the trade-off and a failure mode, not only the definition.
Do Amazon or Infosys ask these computer vision questions?
This page cites no public source that a named company asks any of these questions, so it makes no such claim. Treat the list as the topics worth knowing for any computer vision role.
What is a computer vision system design interview?
A computer vision system design interview is a round where you design a whole vision system, such as data collection, model choice, metrics and serving, rather than answer one question. This page covers the building blocks; the data, the metric and the latency budget are the parts to prepare.
Is there a downloadable list of computer vision interview questions and answers?
This page is the list, kept on the web so each answer links to its source. For wider neural-network practice, use the deep learning interview questions.
Which computer vision projects should you show?
Show one project where you can explain the data, the metric you chose and one failure you found. That lets you answer follow-ups with evidence instead of definitions.
- Output size is (W − F + 2P)/S + 1, and a non-integer result means the setup is impossible.
- Parameter sharing cuts the first-layer example from 105,705,600 parameters to 34,944.
- Deeper networks are harder to train, and residual learning is the fix the ResNet paper proposes.
- A pure transformer on image patches can do well, so convolutions are not strictly required.
- COCO averages AP over 10 IoU thresholds and 80 categories, which rewards better localization.
Sources
- Convolutional Neural Networks (CNNs / ConvNets), CS231n course notes — Stanford University (CS231n), 26 Oct 2026
- Deep Residual Learning for Image Recognition (arXiv 1512.03385) — arXiv (He, Zhang, Ren, Sun), 10 Dec 2015
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (arXiv 2010.11929) — arXiv (Dosovitskiy et al.), 22 Oct 2020
- COCO Detection Evaluation — COCO Consortium, 26 Oct 2026
- Mask R-CNN (arXiv 1703.06870) — arXiv (He, Gkioxari, Dollar, Girshick), 20 Mar 2017
- Scaling Vision with Sparse Mixture of Experts (arXiv 2106.05974) — arXiv (Riquelme et al.), 10 Jun 2021
