引導式測驗 · 中等 等級

視覺基準測試中的標籤錯誤

檢查測試標籤錯誤、缺失框、裁決、指標和排名變化、基準版本控制以及擁擠圖像中的歧義。

問題 1 的 8

A test image’s recorded category is wrong. What can happen to a correct model prediction?

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 test image’s recorded category is wrong. What can happen to a correct model prediction?
    • The evaluator automatically corrects the target
    • It may be scored as an error against the flawed label
    • The model necessarily changes its prediction
    • The wrong label cannot alter any metric
  2. A detector finds a real object that has no annotation box. How can the benchmark miscount it?
    • As a true positive despite no matching box
    • As a false negative for the same detection
    • As an apparent false positive
    • As a corrected category without review
  3. A model disagrees with a test label. Which next step best avoids circular reasoning?
    • Trust the highest-confidence model prediction
    • Change the label whenever the current winner disagrees
    • Delete all disputed images without a rule
    • Have independent reviewers adjudicate the image under documented rules
  4. Why version corrected benchmark labels?
    • Results can be tied to a specific rule and dataset release
    • To make model weights improve automatically
    • To guarantee old and new results are identical
    • To remove the need to record annotation rules
  5. Two objects are clearly visible but the task forces one class. What should maintainers inspect first?
    • Whether the model confidence is highest for one object
    • Whether the annotation rule actually defines which object is the target
    • Whether the image file format favors one class
    • Whether one object has brighter colors
  6. A ranking reverses after independently confirmed label corrections. What does that show?
    • The unchanged model became worse in deployment
    • The corrected set covers all environments
    • The model weights changed during rescoring
    • The original comparison was sensitive to test-label quality
  7. What distinction matters when interpreting label errors?
    • Training errors can affect learning; test errors directly affect reported evaluation
    • Training labels are used only after deployment
    • Test labels cannot be erroneous
    • Wrong test labels cannot change score counts
  8. What claim cannot be made from a higher score after fixing a test file alone?
    • That scores used a revised label version
    • That particular reference targets were corrected
    • That the unchanged model became better in real-world use
    • That earlier results used an older benchmark version