Cropping images in Computer Vision - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

What is the shape of the cropped image?

Given a 100x100 RGB image represented as a NumPy array, what will be the shape of the cropped image after executing the following code?

cropped_img = img[20:70, 30:80]

Computer Vision

import numpy as np
img = np.zeros((100, 100, 3))
cropped_img = img[20:70, 30:80]

A (51, 51, 3)

B (50, 50, 3)

C (49, 49, 3)

D (50, 50)

💡 Hint

Remember that slicing in Python includes the start index but excludes the end index.

The slice 20:70 includes indices from 20 up to 69, which is 50 pixels. Similarly, 30:80 includes 50 pixels. The third dimension (color channels) remains 3.

❓ Model Choice

intermediate

Choosing the best cropping method for data augmentation

You want to augment your image dataset by cropping random patches of size 64x64 from 256x256 images during training. Which method is best to implement this in a deep learning pipeline?

A Randomly select top-left coordinates and crop 64x64 patches on the fly during training.

B Use a fixed crop of the center 64x64 pixels for all images.

C Resize images to 64x64 instead of cropping.

D Crop the bottom-right 64x64 pixels for all images.

💡 Hint

Data augmentation should introduce variety to help the model generalize.

Random cropping during training creates diverse inputs, improving model robustness. Fixed crops or resizing reduce variability.

❓ Metrics

advanced

Effect of cropping on model accuracy

You train two image classifiers on the same dataset. Model A uses full 128x128 images. Model B uses center-cropped 64x64 images. After training, Model A achieves 92% accuracy, Model B achieves 85%. What is the most likely reason for the accuracy drop in Model B?

A Model B used a different optimizer causing lower accuracy.

B Center cropping introduced noise that confused the model.

C Model B was trained with fewer epochs than Model A.

D Cropping reduced the amount of useful information available to the model.

💡 Hint

Think about what information might be lost when cropping images smaller.

Cropping to smaller patches removes parts of the image that might contain important features, reducing model performance.

🔧 Debug

advanced

Why does this cropping code raise an error?

Consider this code snippet:

cropped = img[50:30, 20:70]

Why does this raise an error or produce an unexpected result?

Computer Vision

import numpy as np
img = np.zeros((100, 100, 3))
cropped = img[50:30, 20:70]

A The start index is greater than the end index in the first slice, resulting in an empty array.

B The slice syntax is invalid and causes a SyntaxError.

C The image array does not have enough dimensions for this slicing.

D The indices are out of bounds causing an IndexError.

💡 Hint

Check how Python slicing behaves when start > end.

Python slicing returns an empty array if the start index is greater than the end index; it does not raise an error.

🧠 Conceptual

expert

Impact of cropping on convolutional neural network feature maps

When you crop input images before feeding them into a convolutional neural network (CNN), how does this affect the spatial dimensions of the feature maps in the early convolutional layers compared to using full images?

A Feature maps will be larger because cropping increases feature resolution.

B Feature maps will have the same spatial dimensions regardless of input size due to padding.

C Feature maps will have smaller spatial dimensions proportional to the cropped input size.

D Feature maps will be one-dimensional vectors after cropping.

💡 Hint

Consider how input size affects convolution output size.

Convolutional layers produce feature maps whose spatial size depends on input size, kernel size, stride, and padding. Smaller inputs produce smaller feature maps.