引導式測驗 · 中等 等級

影像資料集中的地理偏差測驗

測試地理覆蓋範圍、家庭物品變化、區域與類別效應、本地驗證、標籤上下文和負責任的聲明。

問題 1 的 8

A household-item model misses unfamiliar cookware in another region. Which dataset issue could explain it?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. A household-item model misses unfamiliar cookware in another region. Which dataset issue could explain it?
    • Training images underrepresented that object style and setting
    • Object function has no visual variants
    • All camera sensors are identical across regions
    • Any model must have the same accuracy everywhere
  2. What was the scope of the de Vries and colleagues study?
    • Proving one country’s objects are inherently harder
    • Evaluating object-recognition systems on geographically diverse household-item images
    • Measuring every deployed recognition system
    • Inferring a photographed individual’s nationality
  3. Why can a worldwide average score conceal geographic bias?
    • An average always equals every region’s score
    • Regional labels remove model error automatically
    • Every region has an identical object distribution
    • Errors can differ sharply across regions while averaging together
  4. A regional score is lower, but that region contains different object classes. What comparison helps interpret the gap?
    • Treat the lower raw score as proof of geography as cause
    • Drop rare classes without disclosing it
    • Compare performance for the same classes across regions
    • Use only the original training-region accuracy
  5. Why is counting images by country alone insufficient?
    • A country count verifies category coverage
    • More photos guarantee local label quality
    • It can hide repeated products, collection channels and missing local categories
    • Each country has a single uniform visual setting
  6. Which test best supports a proposed local deployment?
    • An independent sample of locally relevant objects and camera conditions
    • Only training photos from the original region
    • Only the worldwide average without local slices
    • Only near-duplicate images from one store
  7. A model fails more often in one region. Which interpretation avoids blaming users?
    • Assume residents present objects incorrectly
    • Treat nationality as a visible object feature
    • Hide the regional results from users
    • Investigate data, labels, class mix and capture conditions
  8. Which claim is warranted after a model passes one geographically diverse benchmark?
    • Worldwide reliability is permanently established
    • It worked on that benchmark, while untested places and objects remain uncertain
    • Local category names must already be correct
    • New products require no further evaluation