R-CNN family overview in Computer Vision - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding the main difference between R-CNN and Fast R-CNN

What is the key difference between the original R-CNN and Fast R-CNN in how they process images for object detection?

A R-CNN uses a single CNN for the whole image, while Fast R-CNN uses multiple CNNs for each region proposal.

B Fast R-CNN does not use region proposals, unlike R-CNN which relies on them.

C Fast R-CNN processes the entire image with a single CNN forward pass, while R-CNN processes each region proposal separately.

D R-CNN uses a fully convolutional network, while Fast R-CNN uses a fully connected network.

💡 Hint

Think about how many times the CNN runs on the image in each method.

R-CNN runs the CNN separately on each region proposal, which is slow. Fast R-CNN runs the CNN once on the whole image and then extracts features for each region, making it faster.

❓ Model Choice

intermediate

Choosing the right R-CNN variant for real-time detection

You want to build a real-time object detection system on a mobile device with limited computing power. Which R-CNN family model is the best choice?

A Faster R-CNN

B Fast R-CNN

C R-CNN

D Mask R-CNN

💡 Hint

Consider which model introduced a faster region proposal method.

Faster R-CNN introduced a Region Proposal Network (RPN) that speeds up detection, making it more suitable for real-time applications compared to R-CNN and Fast R-CNN.

❓ Predict Output

advanced

Output shape of feature map in Faster R-CNN backbone

Given an input image of size 224x224x3 passed through a backbone CNN that reduces spatial dimensions by a factor of 16, what is the shape of the output feature map?

Computer Vision

import torch
import torch.nn as nn

input_tensor = torch.randn(1, 3, 224, 224)

class SimpleBackbone(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv = nn.Conv2d(3, 256, kernel_size=3, stride=2, padding=1)
        self.pool = nn.MaxPool2d(2)
    def forward(self, x):
        x = self.conv(x)  # halves size
        x = self.pool(x)  # halves size again
        x = self.pool(x)  # halves size again
        return x

model = SimpleBackbone()
output = model(input_tensor)
output.shape

A torch.Size([1, 256, 56, 56])

B torch.Size([1, 256, 7, 7])

C torch.Size([1, 256, 14, 14])

D torch.Size([1, 256, 28, 28])

💡 Hint

Each stride 2 or pooling halves the spatial size.

Starting from 224x224, conv with stride 2 halves to 112x112, then first pool halves to 56x56, second pool halves to 28x28. So output shape is (1, 256, 28, 28).

But the code applies pooling twice after conv, so final size is 28x28.

❓ Metrics

advanced

Evaluating Mask R-CNN segmentation output

Mask R-CNN outputs a mask for each detected object. Which metric best measures how well the predicted mask matches the true mask?

A Intersection over Union (IoU)

B Accuracy

C Mean Squared Error (MSE)

D Precision

💡 Hint

Think about overlap between predicted and true masks.

IoU measures the overlap between predicted and true masks, making it the best metric for segmentation quality.

🔧 Debug

expert

Debugging slow training in Faster R-CNN

You notice that training Faster R-CNN on your dataset is very slow. You suspect the bottleneck is in the region proposal network (RPN). Which of the following changes will most likely speed up training without hurting accuracy?

A Increase the number of anchors per location in the RPN.

B Reduce the input image size to a smaller resolution.

C Remove the RPN and use external region proposals.

D Increase the number of training epochs.

💡 Hint

Think about how input size affects computation.

Reducing input image size reduces computation in the backbone and RPN, speeding up training without necessarily hurting accuracy much.

Increasing anchors or epochs increases computation. Removing RPN loses the speed advantage of Faster R-CNN.