Chapter 15 Quiz: Computer Vision for Business

Chapter 15 Quiz: Computer Vision for Business

Multiple Choice

Question 1. A standard color digital image with a resolution of 1920 x 1080 pixels contains approximately how many individual data values?

(a) 2 million
(b) 4 million
(c) 6 million
(d) 8 million

Question 2. In a convolutional neural network, what is the primary function of convolutional filters (kernels)?

(a) To reduce the size of the image before processing
(b) To detect spatial patterns (edges, textures, shapes) in the input data
(c) To convert color images to grayscale for faster processing
(d) To classify the image into predefined categories

Question 3. Which of the following best describes the hierarchical feature learning in a CNN?

(a) Each layer independently analyzes the entire image from scratch
(b) Early layers detect complex objects and later layers detect simple edges
(c) Early layers detect simple patterns (edges, gradients) and deeper layers combine these into increasingly complex features (shapes, object parts, objects)
(d) All layers detect the same types of features at different scales

Question 4. What is the primary purpose of pooling layers in a CNN?

(a) To increase the resolution of feature maps for more detailed analysis
(b) To reduce spatial dimensions, decrease computational cost, and provide a degree of translation invariance
(c) To add color information to grayscale feature maps
(d) To connect the convolutional layers to the output classification layer

Question 5. A computer vision model examines a retail shelf image and produces the following output: "Product: Cheerios, Confidence: 0.94, Location: (x=120, y=340, w=85, h=110)." Which computer vision task produced this output?

(a) Image classification
(b) Object detection
(c) Semantic segmentation
(d) Image generation

Question 6. What distinguishes instance segmentation from semantic segmentation?

(a) Instance segmentation is faster but less accurate
(b) Semantic segmentation classifies every pixel, while instance segmentation only processes selected regions
(c) Instance segmentation classifies every pixel AND distinguishes individual objects of the same class, while semantic segmentation treats all objects of the same class as one entity
(d) Instance segmentation works on video while semantic segmentation works only on still images

Question 7. Transfer learning for computer vision typically involves:

(a) Training a model from scratch on a very large, domain-specific dataset
(b) Using a model pre-trained on a large general dataset (like ImageNet) and adapting it to a domain-specific task with less data
(c) Transferring images from one storage system to another for processing
(d) Using the same model architecture for multiple different tasks without any modification

Question 8. According to the chapter, what was the largest cost component in developing Athena's custom shelf analytics model?

(a) GPU compute for model training
(b) Data labeling (annotating products with bounding boxes)
(c) Cloud API usage for inference
(d) Camera hardware installation

Question 9. Which of the following is NOT an advantage of edge deployment for computer vision?

(a) Lower latency (faster inference times)
(b) Reduced bandwidth requirements
(c) Easier model updates and retraining
(d) Enhanced privacy (images never leave the premises)

Question 10. Athena's shelf analytics pilot in 50 stores achieved which of the following results?

(a) 50% reduction in lost sales, representing $17.5M in savings
(b) 12% reduction in lost sales from out-of-stock items, representing $4.2M in annualized savings
(c) 30% reduction in inventory holding costs, representing $8.1M in savings
(d) 25% improvement in cashier efficiency, representing $3.6M in labor savings

Question 11. When should a business prefer a cloud vision API over a custom-trained model?

(a) When domain-specific accuracy above 95% is required
(b) When the task involves standard objects and scenes, volume is moderate, and 80-90% accuracy is acceptable
(c) When images contain sensitive data that cannot leave the corporate network
(d) When real-time processing with sub-100ms latency is required

Question 12. Which of the following was a key finding of Joy Buolamwini and Timnit Gebru's 2018 research on facial recognition systems?

(a) Facial recognition systems were equally accurate across all demographic groups
(b) Commercial facial recognition systems had significantly higher error rates for darker-skinned women compared to lighter-skinned men
(c) Open-source models outperformed commercial systems in all categories
(d) Facial recognition accuracy depends solely on image resolution, not demographic characteristics

Question 13. A food manufacturer is considering computer vision for quality inspection on a production line that moves at 120 items per minute. Which deployment architecture is most appropriate?

(a) Cloud-only processing via API, since cloud servers have more computing power
(b) Edge deployment with a model optimized for real-time inference, since latency requirements preclude cloud round-trips
(c) Manual inspection augmented by periodic cloud-based batch analysis
(d) No CV solution is feasible at this line speed

Question 14. Which model optimization technique reduces the numerical precision of model weights (e.g., from 32-bit to 8-bit) to decrease model size with minimal accuracy loss?

(a) Pruning
(b) Knowledge distillation
(c) Quantization
(d) Regularization

Question 15. According to the chapter, Athena's computer vision governance policy includes all of the following EXCEPT:

(a) No facial recognition capability in the shelf analytics system
(b) Employee union audit rights over the CV system
(c) Automated employee performance scoring based on camera data
(d) 30-day raw image retention with automatic deletion

Short Answer

Question 16. Explain why the chapter recommends "starting with cloud APIs" when evaluating a computer vision opportunity. What information does an API proof-of-concept provide that helps a team decide whether to build a custom model?

Question 17. Describe Athena's hybrid edge-cloud architecture for shelf analytics. What specific benefits did this architecture provide compared to a pure cloud approach?

Question 18. A retail company is considering deploying overhead cameras for anonymous foot traffic analysis (counting people, tracking movement patterns, no facial recognition). Identify three ethical considerations the company should address before deployment, and for each, propose a specific governance measure.

Question 19. The chapter states that in computer vision projects, "data labeling typically accounts for 80-90 percent of the total development cost." Explain why labeling is so expensive for CV applications compared to structured data projects, and describe two strategies for reducing labeling costs.

Question 20. Compare the regulatory environment for computer vision in healthcare (medical imaging) versus retail (shelf analytics). Why does the same underlying technology face dramatically different regulatory burdens depending on the application domain?

Scenario-Based Questions

Question 21. An insurance company wants to use computer vision to estimate vehicle damage from photos submitted by policyholders. The system should identify the damaged parts (hood, bumper, door panel, etc.), estimate the severity of damage, and produce a preliminary cost estimate.

(a) Which computer vision task(s) would this system require? (b) Would you recommend a cloud API, a custom model, or a hybrid approach? Justify your answer. (c) Identify two risks specific to this application that must be addressed before deployment.

Question 22. A pharmaceutical company discovers that its pill inspection system (trained to detect cracked, discolored, or misshapen tablets) performs well on white tablets but has a 40% higher miss rate on colored tablets (blue, red, pink). The system was trained primarily on images of white tablets because 70% of the company's products are white.

(a) What type of bias does this represent? (b) What is the business risk of this bias? (c) Propose a remediation plan.

Question 23. Ravi is presenting the shelf analytics expansion plan to Athena's board. A board member asks: "Why don't we add facial recognition so we can identify loyal customers when they walk in and offer personalized greetings and promotions? Amazon and other tech companies do this."

Draft Ravi's response. Address the business case, the ethical concerns, the regulatory risks, and the potential impact on customer trust. Reference specific examples and regulatory frameworks from the chapter.

Answer Key

(c) — 1920 x 1080 x 3 channels = 6,220,800 values, approximately 6 million.
(b) — Convolutional filters detect spatial patterns in input data.
(c) — CNNs build hierarchical representations from simple to complex features.
(b) — Pooling reduces dimensions, computation, and provides translation invariance.
(b) — Object detection produces class labels with confidence scores and bounding box coordinates.
(c) — Instance segmentation classifies pixels and distinguishes individual object instances.
(b) — Transfer learning adapts pre-trained models to new tasks with less data.
(b) — Data labeling at $2.40 per image ($36,000 total) far exceeded training compute ($12).
(c) — Edge deployment makes model updates harder since models must be pushed to distributed devices.
(b) — The pilot achieved a 12% reduction in OOS-related lost sales, saving $4.2M annualized.
(b) — Cloud APIs are appropriate for standard tasks at moderate volume with acceptable accuracy.
(b) — The Gender Shades study found error rates up to 34.7% for darker-skinned women vs. 0.8% for lighter-skinned men.
(b) — At 120 items per minute (0.5 seconds per item), cloud latency is too high; edge deployment is required.
(c) — Quantization reduces numerical precision of weights.
(c) — Athena's policy explicitly prohibits using CV data for employee performance evaluation.
See exercises section for extended answers.
See exercises section for extended answers.
See exercises section for extended answers.
See exercises section for extended answers.
See exercises section for extended answers. 21-23. See Appendix B: Answers to Selected Exercises.