Depth estimation basics in Computer Vision - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

What does depth estimation predict?

In simple terms, what is the main output of a depth estimation model when given a single image?

A A map showing the distance of each pixel from the camera

B A classification label for the type of scene in the image

C A color-enhanced version of the original image

D A set of bounding boxes around detected objects

💡 Hint

Think about what 'depth' means in a photo.

Depth estimation models predict how far each point in the image is from the camera, producing a depth map.

❓ Predict Output

intermediate

Output shape of depth map from CNN

Given a convolutional neural network that takes a 256x256 RGB image and outputs a depth map, what is the most likely shape of the output tensor?

Computer Vision

import torch
import torch.nn as nn

class SimpleDepthNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv = nn.Conv2d(3, 1, kernel_size=3, padding=1)
    def forward(self, x):
        return self.conv(x)

model = SimpleDepthNet()
input_tensor = torch.randn(1, 3, 256, 256)
output = model(input_tensor)
output.shape

A torch.Size([1, 3, 256, 256])

B torch.Size([1, 1, 254, 254])

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

D torch.Size([256, 256])

💡 Hint

Check the padding and output channels of the convolution.

The Conv2d layer outputs 1 channel with the same height and width due to padding=1.

❓ Model Choice

advanced

Best model type for monocular depth estimation

Which model architecture is most suitable for estimating depth from a single RGB image?

A A simple fully connected network without convolution

B A convolutional encoder-decoder network with skip connections

C A recurrent neural network designed for text sequences

D A clustering algorithm like K-means

💡 Hint

Think about preserving spatial details and image features.

Encoder-decoder CNNs with skip connections capture image features and reconstruct detailed depth maps.

❓ Metrics

advanced

Choosing the right metric for depth estimation

Which metric is commonly used to measure the accuracy of predicted depth maps compared to ground truth?

A Mean Absolute Error (MAE) between predicted and true depth values

B Intersection over Union (IoU) for object detection

C Accuracy of classification labels

D BLEU score for language translation

💡 Hint

The metric should measure how close predicted distances are to actual distances.

MAE measures the average absolute difference between predicted and true depth values, making it suitable for depth estimation.

🔧 Debug

expert

Why does the depth model output all zeros?

You trained a depth estimation model but the output depth map is all zeros for every input image. Which is the most likely cause?

A The optimizer learning rate is too high

B The input images are grayscale instead of RGB

C The batch size is set to 1 during training

D The final activation function is ReLU and the model weights are initialized to zero

💡 Hint

Consider how ReLU and zero weights affect output values.

If weights start at zero and ReLU is used, the output will remain zero because ReLU outputs zero for zero or negative inputs.