引導式測驗 · 中等 等級
細粒度視覺分類測驗
測試微妙的班級線索、CUB 風格的部分註釋、個人級別的分割、預訓練重疊、模糊的視圖和棄權。
問題 1 的 8
How does fine-grained classification differ from a broad bird-versus-car task?
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。
- 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
- 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
- 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
- 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
- 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
- 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
- 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
- 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