Computer Vision Interview Questions

Question 1

In the context of the U-Net architecture, what is the purpose of the skip connections?

Accepted Answer

To concatenate high-resolution spatial features from the encoder with upsampled features in the decoder

Question 2

Why is ImageNet considered a milestone dataset in computer vision?

Accepted Answer

Its massive scale of over a million labeled images enabled the deep learning revolution in CNNs

Question 3

When addressing extreme class imbalance in a vision dataset, what is a primary risk of aggressively oversampling the minority class images?

Accepted Answer

The model may heavily overfit to the exact specific features of the duplicated minority images

Question 4

Why is the Softmax function generally placed at the end of an image classification network?

Accepted Answer

To convert the raw logits into a normalized probability distribution summing to 1

Question 5

What is the defining characteristic of a Depthwise Separable Convolution?

Accepted Answer

It separates the convolution into a spatial channel-by-channel pass followed by a 1x1 cross-channel pass

Question 6

How does applying L2 Regularization (Weight Decay) affect the weights in a Convolutional Neural Network?

Accepted Answer

It encourages the network to keep weight values small by penalizing the sum of their squares

Question 7

Which component of the loss function in YOLO penalizes the model for predicting bounding boxes incorrectly?

Accepted Answer

Sum-Squared Error on the coordinates and dimensions

Question 8

How does the feature representation of an object change as it passes deeper into a standard CNN?

Accepted Answer

The spatial resolution decreases while the semantic complexity and abstraction increase

Question 9

A standard convolutional layer (without pooling) provides which mathematical property with respect to translation?

Accepted Answer

Translation equivariance

Question 10

The Gray World assumption is utilized in computer vision primarily for which preprocessing task?

Accepted Answer

Automatic white balancing and color constancy

Question 11

What is the main limitation of using standard Mean Squared Error (MSE) loss for image generation tasks like autoencoders?

Accepted Answer

It tends to produce blurry images because it averages out plausible sharp variations

Question 12

What characterizes a spatial separable convolution?

Accepted Answer

It breaks a kxk spatial kernel into two sequential 1D kernels of size kx1 and 1xk

Question 13

How does a feature map change when passed through a standard 2x2 Max Pooling layer with a stride of 2?

Accepted Answer

The spatial resolution (width and height) is strictly halved

Question 14

In an autoencoder used for image denoising, why is the bottleneck layer critical?

Accepted Answer

It forces the network to learn a compressed representation of the most important features

Question 15

What phenomenon occurs when a GAN generator starts producing a very limited variety of outputs, regardless of the input noise?

Accepted Answer

Mode Collapse

Question 16

In a convolutional layer, what does a dilation rate of 1 represent?

Accepted Answer

A standard convolution where kernel elements are adjacent with no gaps

Question 17

How does dilated (atrous) convolution increase the receptive field without increasing computational cost?

Accepted Answer

By inserting zeros between the weights of the convolutional filter

Question 18

How does the R-CNN object detection framework initially propose regions to classify?

Accepted Answer

Using selective search to extract region proposals

Question 19

Which of the following is a direct consequence of using 1x1 convolutions in architectures like Inception?

Accepted Answer

They perform depth-wise dimensionality reduction to decrease computational cost

Question 20

Which concept refers to a model ability to maintain performance regardless of an object scale in the image?

Accepted Answer

Scale Invariance

Question 21

If a standard convolution requires N parameters, how many parameters does a grouped convolution require if the channels are divided into G groups?

Accepted Answer

N divided by G

Question 22

In a Variational Autoencoder (VAE), what is the purpose of the Kullback-Leibler (KL) divergence term in the loss function?

Accepted Answer

To ensure the latent space distribution closely matches a standard normal distribution

Question 23

If a CNN layers 3x3 convolutions consecutively, what is the effective receptive field of the second layer with respect to the original input? (Assume stride 1, no dilation) ?

Accepted Answer

5x5

Question 24

In the context of evaluating an object detector, what trade-off occurs when you lower the confidence threshold?

Accepted Answer

Recall increases but Precision likely decreases

Question 25

If an architecture cascades two 3x3 convolutional layers with a stride of 1, what is their combined receptive field on the original input?

Accepted Answer

5x5

Computer Vision Interview Questions

Ready to test yourself?

Want to study the theory first?

Practice with 8,400+ similar questions