Chapter 6 Quiz: Computer Vision

Test your understanding of key concepts from this chapter. Try answering each question before checking the explanation.

Question 1. A standard color digital image stores each pixel as three numbers (red, green, blue), each ranging from 0 to 255. What does a computer vision system initially "see" when it processes this image?

(a) Objects and their relationships (b) A grid of numerical color values (c) Edges, textures, and shapes (d) A compressed version of the scene

Answer: (b). At the most fundamental level, a digital image is nothing more than a grid of numerical values. The entire challenge of computer vision is extracting meaning (objects, relationships, scenes) from these raw numbers. Options (c) and (a) represent progressively higher levels of the vision hierarchy that the system must learn to extract.

Question 2. What is the key innovation of convolutional neural networks (CNNs) compared to earlier approaches to computer vision?

(a) CNNs use hand-written rules to identify specific objects (b) CNNs process images all at once rather than in pieces (c) CNNs learn which visual features to detect from training data rather than having features programmed in (d) CNNs only work on grayscale images, which makes them faster

Answer: (c). The breakthrough of CNNs is that they learn what features (edges, textures, shapes) to look for during training, rather than relying on features manually specified by programmers. Earlier "classical" computer vision relied on hand-coded rules about edges and shapes, which worked in controlled settings but failed in the messy real world.

Question 3. A CNN trained to recognize animals processes an image. The first layer detects edges; the second layer detects textures and shapes; the third layer detects body parts; the final layer identifies the animal. This progression illustrates:

(a) The adversarial nature of neural networks (b) The hierarchical feature learning of CNNs (c) The difference between classification and detection (d) The principle of image segmentation

Answer: (b). CNNs learn hierarchical features — simple features (edges) in early layers combine into increasingly complex features (textures, parts, objects) in later layers. This mirrors the hierarchy of visual processing: pixels, edges, textures, parts, objects, scenes.

Question 4. What is the primary difference between object detection and image segmentation?

(a) Object detection works on video; segmentation works only on still images (b) Object detection identifies objects and their approximate locations; segmentation labels every pixel in the image (c) Object detection requires a CNN; segmentation does not (d) Object detection is more accurate than segmentation

Answer: (b). Object detection identifies objects and draws bounding boxes around them. Segmentation assigns a label to every single pixel, providing precise outlines rather than approximate boxes. Both can work on still images or video, and both typically use neural network architectures.

Question 5. The Gender Shades study by Buolamwini and Gebru found that commercial facial recognition systems had the highest error rates for:

(a) Lighter-skinned male faces (b) Darker-skinned male faces (c) Lighter-skinned female faces (d) Darker-skinned female faces

Answer: (d). The study found error rates as high as 34.7% for darker-skinned female faces, compared to below 1% for lighter-skinned male faces. This disparity was traced to training datasets that overrepresented lighter-skinned faces, particularly male faces.

Question 6. Which of the following best describes a "face embedding"?

(a) A photograph of a face stored in a database (b) A numerical representation of a face's unique geometric features (c) The physical process of scanning a face with a camera (d) A face that has been modified by a deepfake algorithm

Answer: (b). A face embedding is a mathematical representation — essentially a list of numbers — that captures the unique geometry of a face (distances between features, proportions, angles). Two photos of the same person produce similar embeddings; different people produce different embeddings. The system compares these numerical representations, not the photos themselves.

Question 7. What is the "liar's dividend" in the context of deepfakes?

(a) The profit deepfake creators earn from their content (b) The advantage that deepfake detection tools have over generators (c) The ability of people to dismiss real evidence as fake because deepfakes exist (d) The cost savings from using AI-generated actors instead of real ones

Answer: (c). The "liar's dividend" refers to the perverse benefit that the mere existence of deepfake technology provides to people who want to deny the authenticity of real evidence. Once people know convincing fakes are possible, any inconvenient video or audio can be dismissed with "that's a deepfake" — even when it's genuine.

Question 8. An adversarial example is:

(a) A real-world object that is difficult for any vision system to classify (b) An image with carefully calculated, often imperceptible modifications designed to fool a neural network (c) A low-quality image that any vision system would struggle with (d) An image that looks fake to humans but real to AI

Answer: (b). Adversarial examples involve deliberate, calculated modifications to images — often imperceptible to the human eye — that cause neural networks to misclassify them with high confidence. A famous example is adding carefully computed noise to a panda image, causing a CNN to classify it as a gibbon with 99% confidence while appearing unchanged to human observers.

Question 9. A self-driving car's vision system correctly identifies 99.9% of pedestrians in test conditions. Why might this not be sufficient for real-world deployment?

(a) 99.9% accuracy is below the threshold for any commercial deployment (b) The 0.1% failure rate could involve edge cases with catastrophic consequences, and the test conditions may not represent the full range of real-world scenarios (c) Self-driving cars don't actually use computer vision (d) Pedestrians don't need to be identified, only vehicles do

Answer: (b). A 0.1% failure rate sounds impressive, but on a road with thousands of pedestrian encounters, it means regular failures. More importantly, the failures are not random — they cluster in edge cases (unusual lighting, unusual pedestrian appearance, unusual behavior) that may not be well-represented in test data. A single failure in a real-world driving scenario can be fatal.

Question 10. Which recurring theme from the book is most directly illustrated by the fact that MedAssist AI performs better on patient populations similar to its training data?

(a) AI literacy as a civic skill (b) Durable frameworks (c) Tools built by humans — AI systems carry human biases, incentives, and blind spots (d) Capability vs. understanding

Answer: (c). MedAssist AI's performance disparity directly illustrates that AI systems are tools built by humans. The training data was collected from specific hospitals serving specific populations, and those human choices shaped the system's capabilities and limitations. The biases in the data became biases in the system.

Bonus Question. Consider the following scenario: A university installs a computer vision system that uses facial recognition to track class attendance. The system also uses object detection to monitor whether students are using phones during lectures. Apply two concepts from this chapter to evaluate this proposal.

Sample answer: First, accuracy disparities (from the Gender Shades research) suggest the attendance system would likely perform unevenly across student demographics, potentially marking some students absent when they are present. This could disproportionately affect students of color. Second, the consent problem (from Section 6.4) applies — students in required courses cannot meaningfully opt out of biometric surveillance, creating a coercive dynamic where attending class means submitting to facial recognition. Additional concepts that could apply: chilling effects on classroom participation, function creep if data is used for purposes beyond attendance, and the human-in-the-loop question of whether faculty should override system decisions.