Computer vision interview questions and answers
By Vikram · Updated 26 Oct 2026
Computer vision interviews check that you understand the maths of convolutions, how detection and segmentation models are trained and evaluated, and when to reach for transfer learning or a Vision Transformer. Here are the questions to expect.
Images and convolutions
- Image as an array: height × width × channels. OpenCV loads colour images as BGR, not RGB.
- Convolution output size: (N − K + 2P) / S + 1. A 32 × 32 input with a 5 × 5 kernel, no padding, stride 1 gives 28 × 28.
- Parameters of a conv layer: K × K × C_in × C_out + C_out biases. A 3 × 3 conv from 64 to 128 channels has 73,856.
- Pooling: 2 × 2 max pooling with stride 2 halves height and width and keeps the strongest response.
- Receptive field: two stacked 3 × 3 convolutions see 5 × 5, with fewer parameters than one 5 × 5.
- 1 × 1 convolutions: mix channels cheaply and change the channel count (bottlenecks).
Detection and segmentation
- Classification vs detection vs segmentation: one label per image, a box and label per object, a label per pixel.
- IoU: intersection over union of two boxes; a prediction usually counts as correct above 0.5.
- NMS (non-maximum suppression): keeps the best box and removes overlapping duplicates.
- mAP@0.5:0.95: mean average precision averaged over IoU thresholds from 0.5 to 0.95: the COCO standard.
- One-stage vs two-stage detectors: YOLO and SSD predict boxes in one pass (fast); Faster R-CNN proposes regions first (often more accurate).
- Semantic vs instance segmentation: all “person” pixels share one label, or each person gets their own mask (Mask R-CNN).
Modern models
- Transfer learning: start from an ImageNet-pretrained backbone and fine-tune; you need far less data.
- Data augmentation: flips, crops and colour jitter teach invariance and reduce overfitting.
- Vision Transformer (ViT): splits the image into patches (e.g. 16 × 16), embeds each as a token and runs a Transformer.
- CLIP: trains image and text encoders contrastively on image–caption pairs, which enables zero-shot classification.
Tip: Check the maths daily with the computer vision questions with answers from the Daily Quiz, and build a vision project at an AI hackathon.
Frequently asked questions
How do you calculate the output size of a convolution?
Output = (N − K + 2P) / S + 1, where N is the input size, K the kernel size, P the padding and S the stride.
What is IoU in object detection?
Intersection over union: the overlap area of two boxes divided by the area they cover together. It measures how well a predicted box matches the true one.
What does non-maximum suppression do?
It removes duplicate boxes for the same object by keeping the highest-scoring box and discarding others that overlap it beyond an IoU threshold.
What is a Vision Transformer?
A model that cuts an image into fixed-size patches, turns each into a token and processes them with a Transformer instead of convolutions.