Guide-linked quiz · Medium Level

Fine-Grained Visual Classification Quiz

Test subtle class cues, CUB-style part annotations, individual-level splits, pretraining overlap, ambiguous views, and abstention.

Question 1 of 8

How does fine-grained classification differ from a broad bird-versus-car task?

Preview every question8 questions

Answers and explanations appear as you take the quiz.

  1. How does fine-grained classification differ from a broad bird-versus-car task?
    • It only traces pixel boundaries
    • It separates closely related categories with subtle visual differences
    • It requires no class ontology
    • It identifies an individual by default
  2. Why might a second view help distinguish two similar bird species?
    • The second view changes the true category
    • A new angle guarantees a correct model label
    • A diagnostic marking or part may be hidden in the first image
    • Part visibility never affects fine-level labels
  3. Which evaluation split avoids easy recognition of the same specimen?
    • Keep all photos of one individual in one partition
    • Randomly split adjacent views of the same item
    • Place one view of each specimen in train and another in test
    • Choose test individuals after seeing model predictions
  4. Why does the CUB-200-2011 site caution about pretrained networks?
    • Bird images cannot be classified after pretraining
    • All part annotations are invalid
    • Pretraining always reduces fine-grained accuracy
    • Pretraining data may overlap benchmark test images
  5. A model predicts a species from a blurred photo with a hidden beak. What is the main concern?
    • Blur makes every species biologically identical
    • Its score may exceed the evidence visible for that distinction
    • A high score proves a hidden diagnostic cue was visible
    • Background alone determines the species label
  6. Which report shows where a fine-grained classifier struggles?
    • Only total dataset bytes
    • Only average training loss
    • Only the number of network layers
    • Confusion counts between particular related classes
  7. A product variant differs only by a serial mark outside the photo. What should the system say?
    • The available image cannot establish that variant reliably
    • Broad object shape reveals the serial variant
    • The largest class score supplies the missing mark
    • Surface texture always identifies the exact variant
  8. When is an abstain or expert-review route useful?
    • Only when the image is already unambiguous
    • Whenever two categories are easy to separate
    • When needed diagnostic features are hidden or category labels remain ambiguous
    • Only when no pretrained weights are used

Keep testing yourself

More quizzes picked for your level.

View all quizzes