Computer vision is transforming industries — from autonomous vehicles and medical imaging to retail analytics and industrial quality control. Computer vision engineers are in high demand, and interviews test both theoretical understanding of the architectures and practical knowledge of training, evaluating, and deploying vision models. This guide covers the 35 most important computer vision interview questions with detailed answers for 2026.
Image Classification
Q1. What is image classification and what are the main CNN architectures used?
Image classification assigns a label from a predefined set of categories to an input image. It is the foundational computer vision task. Key architectures: LeNet (1998) — first practical CNN, used for digit recognition. AlexNet (2012) — won ImageNet with top-5 error of 15.3% vs 26.2% for prior methods, launching the deep learning era. VGG-16/19 (2014) — very deep networks (16-19 layers) using only 3×3 convolutions, still widely used as a backbone. ResNet (2015) — residual connections solve vanishing gradients, enabling 50-200 layer networks. ResNet-50 is still the most common backbone for transfer learning. Inception/GoogLeNet — inception modules apply multiple filter sizes in parallel for efficiency. EfficientNet (2019) — systematically scales width, depth, and resolution using neural architecture search, achieving state-of-the-art accuracy at lower compute. Vision Transformer (ViT, 2020) — applies self-attention directly to image patches, now dominant for large-scale pretraining.
Q2. What is transfer learning in computer vision and when should you use it?
Transfer learning uses weights pretrained on a large dataset (typically ImageNet — 1.2M images, 1000 classes) as the starting point for a new task. The pretrained network has learned to detect edges, textures, shapes, and complex visual patterns from ImageNet. These features transfer remarkably well to new visual domains. When to use: almost always — unless your domain is very different from natural images (medical histology, satellite imagery, X-rays) or you have millions of labelled images. Two strategies: Feature extraction — freeze all pretrained layers, add and train only a new classification head. Fast, low compute, works well with < 1,000 labelled examples. Fine-tuning — unfreeze some or all pretrained layers and train with a very small learning rate (1e-4 to 1e-5). More compute, better results when you have 10,000+ labelled examples. Tip: always freeze BatchNorm layers when fine-tuning — their running statistics (trained on ImageNet) should not be overwritten by your small dataset.
Q3. What is data augmentation and what techniques are most effective for images?
Data augmentation artificially increases training set size and diversity by applying random transformations, teaching the model invariance to irrelevant variations. Basic geometric: horizontal flip (left-right invariant tasks), random crop and resize (positional invariance), random rotation (±30°), random shear, random zoom. Basic photometric: brightness, contrast, saturation, hue jitter (simulates different lighting), random grayscale, Gaussian blur, Gaussian noise. Advanced: Cutout/CoarseDropout — randomly set rectangular patches to zero, forcing the model to use global context not local patches. MixUp — blend two images and their labels linearly (λ × img1 + (1-λ) × img2, same blending for labels). CutMix — cut a patch from one image and paste into another, mix labels proportionally. RandAugment — automatically searches for the optimal combination of augmentations for each dataset. AutoAugment. In medical imaging: elastic deformations, stain normalisation, careful application (flipping a histology image is OK; inverting a chest X-ray is not).
Q4. What is the ImageNet Large Scale Visual Recognition Challenge (ILSVRC)?
ILSVRC was an annual competition (2010-2017) where teams competed to classify images from ImageNet’s 1000-class subset. It was the primary benchmark that drove progress in deep learning for computer vision. AlexNet’s 2012 win (15.3% top-5 error vs 26.2% second place) demonstrated the superiority of deep CNNs and started the modern deep learning era. Subsequent winners: ZFNet (2013), VGG (2014), GoogLeNet (2014, same year), ResNet (2015, 3.6% — surpassing human performance of ~5%), Squeeze-and-Excitation Networks (2017). The dataset: 1.2M training, 50K validation, 150K test images across 1000 categories. Pretrained ImageNet weights are the starting point for virtually all vision transfer learning today.
Object Detection
Q5. What is object detection and how does it differ from image classification?
Image classification outputs one label per image. Object detection outputs multiple bounding boxes, each with a class label and a confidence score — where are the objects and what are they? It combines localisation (finding objects’ positions) and classification (identifying object types). Detection approaches: One-stage detectors (YOLO, SSD, RetinaNet) predict boxes and classes in a single forward pass — fast (real-time capable), slightly lower accuracy. Two-stage detectors (R-CNN family: Faster R-CNN, Mask R-CNN) first generate region proposals, then classify each region — more accurate but slower. YOLO v8 and RT-DETR are state-of-the-art in 2026 for real-time detection. DETR (Detection Transformer) treats detection as a set prediction problem using Transformers — no hand-designed anchors or NMS needed.
Q6. Explain anchor boxes and their role in object detection.
Anchor boxes are predefined bounding boxes of various scales and aspect ratios placed at each location in the feature map. The model predicts offsets from these anchors, rather than absolute box coordinates — the prediction is more stable because anchors provide reasonable starting points. For each anchor, the model predicts: (1) objectness score — does this anchor contain an object? (2) class probabilities — which class? (3) bounding box offsets — δx, δy, δw, δh to adjust the anchor. During training, anchors are matched to ground truth boxes (IoU > 0.5 = positive, IoU < 0.4 = negative). YOLO uses anchors per grid cell. Faster R-CNN generates ~300 anchors per image, classifies each. Anchor-free detectors (FCOS, CornerNet, DETR) predict boxes directly without anchors — simpler but require alternative matching strategies.
Q7. What is IoU (Intersection over Union) and how is it used in detection?
IoU measures the overlap between a predicted bounding box and a ground truth bounding box: IoU = Area(Intersection) / Area(Union). Ranges from 0 (no overlap) to 1 (perfect match). Uses: (1) Matching predictions to ground truth during training — a predicted box with IoU > 0.5 vs a GT box is a true positive. (2) Non-Maximum Suppression (NMS) — after detection, a model produces many overlapping boxes for the same object. NMS suppresses duplicate detections: sort boxes by confidence; select the highest-confidence box; suppress all boxes with IoU > threshold (e.g., 0.5) with the selected box; repeat. (3) Evaluation metric: mean Average Precision (mAP) at IoU threshold 0.5 (mAP@0.5) or averaged over 0.5:0.05:0.95 (COCO metric). mAP@0.5:0.95 is the standard benchmark for comparing detectors.
Q8. What is YOLO and how has it evolved?
YOLO (You Only Look Once, 2015) introduced single-pass object detection: divide the image into an S×S grid; each cell predicts B bounding boxes with confidence scores and C class probabilities; confidence = P(object) × IoU. All predictions are made in one forward pass, enabling real-time detection. YOLO v2 (2016): better anchors from k-means clustering on training boxes, batch normalisation, higher resolution. YOLO v3 (2018): multi-scale predictions (three scales), residual connections. YOLO v4 (2020): CSP (Cross Stage Partial) networks, Mosaic augmentation, PANet neck. YOLO v5 (Ultralytics, 2020): PyTorch implementation, easy-to-use API. YOLO v8 (2023): anchor-free, new architecture, stronger performance, same Ultralytics API. YOLOv8 is the most widely deployed detector in 2026 for production systems requiring real-time performance.
Image Segmentation
Q9. What is the difference between object detection, semantic segmentation, instance segmentation, and panoptic segmentation?
Object detection: bounding boxes + class labels. No pixel-level precision. Semantic segmentation: classify every pixel into a category class. All cars are the same class — two cars share the same “car” label. No distinction between individual instances. Used for autonomous driving scene understanding, satellite imagery analysis. Instance segmentation: combines detection and segmentation — finds each object instance and its precise pixel mask. Two cars = two separate masks. Mask R-CNN is the standard approach. Used for robotics, medical cell counting. Panoptic segmentation: unifies semantic and instance segmentation — every pixel is assigned both a class label and an instance ID. “Stuff” classes (sky, road, vegetation) get semantic labels; “things” classes (cars, people) get instance labels. The most complete scene understanding task.
Q10. What is U-Net and why is it widely used in medical imaging?
U-Net (2015) is an encoder-decoder architecture with skip connections specifically designed for biomedical image segmentation. The encoder downsamples the image through convolutional blocks, learning increasingly abstract features. The decoder upsamples back to the original resolution. The key innovation: skip connections between corresponding encoder and decoder layers. These concatenate low-level spatial detail (from the encoder) with high-level semantic features (from the decoder), enabling precise localisation — critical when segmenting small structures like tumours, cells, or blood vessels. It was designed to work with very few labelled medical images (tens to hundreds), using heavy augmentation to compensate. It remains the dominant architecture for medical image segmentation, with modern variants: U-Net++, Attention U-Net, Swin-UNet (Transformer-based).
Q11. What is DeepLab and what is atrous (dilated) convolution?
DeepLab is a family of semantic segmentation models from Google. Its key innovation is atrous (dilated) convolution — inserting zeros between filter weights to increase the receptive field without increasing parameters or reducing resolution. A 3×3 filter with dilation rate 2 has the receptive field of a 5×5 filter but only 9 parameters. Atrous Spatial Pyramid Pooling (ASPP) applies multiple dilated convolutions with different rates in parallel, capturing multi-scale context. DeepLabv3+ combines ASPP with an encoder-decoder structure for precise boundary segmentation. Dilated convolutions are now used broadly in dense prediction tasks: segmentation, depth estimation, optical flow.
Advanced Topics
Q12. What is the Vision Transformer (ViT) and how does it differ from CNNs?
ViT (2020, Google Brain) applies the Transformer architecture directly to images. The image is divided into fixed-size patches (e.g., 16×16 pixels); each patch is flattened and linearly embedded into a token. A learnable [CLS] token is prepended. Positional embeddings are added. The sequence of tokens is processed by standard Transformer encoder blocks with self-attention. The [CLS] token’s representation is used for classification. Key differences from CNNs: no inductive biases (no assumption of local connectivity or translation invariance — must learn all structure from data); self-attention is global (every patch attends to every other patch from the first layer); requires much more data than CNNs to train from scratch (ViT underperforms ResNet on ImageNet alone, but surpasses it when pretrained on JFT-300M). Modern variants: DeiT (distillation-efficient), Swin Transformer (hierarchical, shifted windows — most practical for dense prediction tasks), BEiT, MAE (masked autoencoding pretraining).
Q13. What is contrastive learning and how is it used in computer vision?
Contrastive learning is a self-supervised approach that learns visual representations without labels by training a model to be similar for different views of the same image (“positive pairs”) and different for views of different images (“negative pairs”). SimCLR (2020): augment each image twice (crop, flip, colour jitter), embed both with an encoder, minimise the cosine distance between same-image pairs (positives) while maximising it against all other pairs in the batch (negatives). MoCo (2020): maintains a memory queue of negative embeddings, enabling large numbers of negatives without large batches. CLIP (2021, OpenAI): contrastive learning between image and text embeddings from 400M image-text pairs — enables zero-shot classification by comparing image embeddings to text description embeddings. DINO (2021): self-distillation with no labels, using teacher-student architecture — produces excellent features for segmentation without any labels.
Q14–21 (Computer vision rapid fire):
Q14. What is batch normalisation vs layer normalisation in vision models? Batch norm normalises across the batch dimension — requires large enough batches; standard in CNNs. Layer norm normalises across the feature dimension — works for batch size of 1; standard in Transformers (ViT). Group norm is a compromise — normalises within groups of channels — works for small batches in CNNs.
Q15. What is the receptive field and why does it matter? The region of the input image that influences a neuron’s output. Deep networks need large receptive fields to detect large objects. Stacking convolutional layers, using pooling, dilated convolutions, or global attention (Transformer) all expand the receptive field.
Q16. What is depthwise separable convolution (MobileNet)? Factorises a standard convolution into depthwise convolution (one filter per input channel) + pointwise convolution (1×1 conv to combine channels). ~8-9x fewer FLOPs for 3×3 conv. MobileNet and EfficientNet use this for lightweight models deployable on mobile/edge devices.
Q17. What is image preprocessing for CNNs? Resize to model’s expected input size (e.g., 224×224), convert to float32, normalise by subtracting ImageNet mean ([0.485, 0.456, 0.406]) and dividing by std ([0.229, 0.224, 0.225]). This aligns input distribution with what the pretrained model expects.
Q18. What is GradCAM? Gradient-weighted Class Activation Mapping — uses the gradient of the class score with respect to the final convolutional feature maps to produce a coarse localisation heatmap showing which regions of the image influenced the prediction. Useful for debugging and explaining CNN decisions.
Q19. What is the difference between precision and recall in object detection? Precision = TP / (TP + FP) — of all detected boxes, what fraction are correct? Recall = TP / (TP + FN) — of all GT objects, what fraction did we detect? Average Precision (AP) is the area under the precision-recall curve. Mean Average Precision (mAP) averages AP across all classes.
Q20. What is optical flow? Estimating the apparent motion of objects between consecutive frames in a video — a 2D vector field where each vector represents the displacement of a pixel from frame t to frame t+1. Used for action recognition, video stabilisation, and autonomous driving. Classic methods: Lucas-Kanade, Horn-Schunck. Deep learning: FlowNet, RAFT.
Q21. What is the difference between 2D and 3D convolution for video? 2D convolution processes spatial dimensions (H×W) only — applied frame by frame, misses temporal patterns. 3D convolution (T×H×W) captures spatio-temporal patterns — used in video classification (C3D, I3D). Computationally expensive. (2+1)D convolution factorises 3D into spatial + temporal: similar accuracy, much cheaper.
Conclusion
Computer vision interviews in 2026 increasingly require knowledge of both CNN fundamentals and Vision Transformers, with practical experience in object detection (YOLO v8) and segmentation (U-Net for medical, Mask R-CNN for general). The highest-value preparation is hands-on: train a fine-tuned ResNet-50 or EfficientNet on a custom dataset using torchvision, implement a YOLO detection pipeline on a real-world dataset, and experiment with a Vision Transformer using Hugging Face transformers. Reading the original papers for ResNet, YOLO, and ViT will answer the “how does it work” questions that appear in every computer vision interview loop.
