U-Net architecture in Computer Vision - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding U-Net's Skip Connections

In the U-Net architecture, what is the main purpose of the skip connections between the encoder and decoder parts?

A To reduce the number of parameters by sharing weights between encoder and decoder layers.

B To combine low-level spatial information from the encoder with high-level features in the decoder to improve segmentation accuracy.

C To increase the depth of the network by adding more convolutional layers in the decoder.

D To perform max pooling operations to downsample the feature maps.

💡 Hint

Think about how the network keeps details from the input image while reconstructing the output.

Skip connections in U-Net pass detailed spatial information from the encoder to the decoder. This helps the decoder recover fine details lost during downsampling, improving segmentation results.

❓ Predict Output

intermediate

Output Shape After U-Net Encoder Block

Given an input tensor of shape (batch_size=1, channels=3, height=128, width=128), what is the output shape after one encoder block in a U-Net that applies two 3x3 convolutions (padding=1, stride=1) followed by a 2x2 max pooling (stride=2)?

Computer Vision

import torch
import torch.nn as nn

class EncoderBlock(nn.Module):
    def __init__(self, in_channels, out_channels):
        super().__init__()
        self.conv = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
            nn.ReLU()
        )
        self.pool = nn.MaxPool2d(kernel_size=2, stride=2)

    def forward(self, x):
        x = self.conv(x)
        p = self.pool(x)
        return x, p

x = torch.randn(1, 3, 128, 128)
block = EncoderBlock(3, 64)
features, pooled = block(x)
output_shape = pooled.shape

A (1, 64, 64, 64)

B (1, 64, 128, 128)

C (1, 3, 64, 64)

D (1, 64, 32, 32)

💡 Hint

Remember that padding keeps spatial size after convolution, and max pooling halves height and width.

Each convolution keeps height and width at 128 due to padding=1. Max pooling with kernel 2 and stride 2 halves height and width to 64. The output channels are 64 as set in the block.

❓ Model Choice

advanced

Choosing the Right Activation for U-Net Output

For a U-Net model designed to perform binary segmentation (classifying each pixel as foreground or background), which activation function is most appropriate to use in the final layer?

A Softmax activation

B ReLU activation

C Tanh activation

D Sigmoid activation

💡 Hint

Consider the output as a probability for each pixel belonging to the foreground class.

Sigmoid activation outputs values between 0 and 1 for each pixel independently, suitable for binary segmentation. Softmax is for multi-class problems, ReLU and Tanh are not probability functions.

❓ Metrics

advanced

Evaluating U-Net Segmentation with Dice Coefficient

Which of the following formulas correctly computes the Dice coefficient for evaluating the overlap between predicted and ground truth segmentation masks?

A Dice = (2 * |Prediction ∩ GroundTruth|) / (|Prediction| + |GroundTruth|)

B Dice = |Prediction ∩ GroundTruth| / |Prediction ∪ GroundTruth|

C Dice = |Prediction ∪ GroundTruth| / (|Prediction| + |GroundTruth|)

D Dice = |Prediction| / |GroundTruth|

💡 Hint

Dice coefficient measures similarity by doubling the intersection over the sum of sizes.

The Dice coefficient formula is 2 times the intersection size divided by the sum of the sizes of prediction and ground truth sets, measuring overlap between masks.

🔧 Debug

expert

Identifying the Cause of Dimension Mismatch in U-Net Decoder

In a U-Net decoder block, a concatenation of the upsampled feature map and the corresponding encoder feature map fails with a dimension mismatch error. Given that the encoder feature map has shape (batch_size, 64, 64, 64) and the upsampled decoder feature map has shape (batch_size, 64, 65, 65), what is the most likely cause?

A The encoder feature map has wrong channel size causing mismatch during concatenation.

B The batch size is different between encoder and decoder feature maps.

C The upsampling operation increased spatial dimensions incorrectly, causing a mismatch with encoder features.

D The concatenation axis is set incorrectly to the batch dimension.

💡 Hint

Check how upsampling changes height and width compared to encoder features.

Upsampling sometimes produces spatial sizes off by one pixel due to rounding. This causes mismatch when concatenating with encoder features of fixed size. Adjusting upsampling or cropping fixes this.