引導式測驗 · 中等 等級
圖像分類說明測驗
測試單標籤和多標籤分類、偵測邊界、遷移學習、分組評估、校準和不確定輸入。
問題 1 的 8
A model labels a whole photo “bicycle.” What does classification alone leave unresolved?
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。
- A model labels a whole photo “bicycle.” What does classification alone leave unresolved?
- Where the bicycle is in the image
- Which label was selected from its vocabulary
- Whether the prediction score can be displayed
- Whether a class label can be assigned to the crop
- A street scene contains a bus and bicycle. Which output setup can represent both as present?
- One mutually exclusive softmax over the scene
- A location box without a multi-label head
- One dominant-category tag only
- Independent multi-label outputs
- Why can a classifier label an unfamiliar object confidently as a known class?
- A high score guarantees a familiar object
- Its output is restricted to learned labels and may not represent “none of these”
- Transfer learning adds every unseen class automatically
- The largest score is validated as an unknown-class detector
- A team has few labeled photos for a new task. Which transfer-learning approach is documented?
- Discard pretrained weights before every small-data task
- Change only display names without fitting a classifier
- Reuse a pretrained feature extractor and train a new final classifier
- Tune only on final test photos
- Why group photos of the same physical item into one partition during evaluation?
- Near duplicates across train and test can inflate the score
- It guarantees each class is equally frequent
- It makes camera calibration unnecessary
- It changes classification into instance segmentation
- A product rejects good items more often than bad ones. Which report helps diagnose this?
- Only the total dataset image count
- Only the architecture layer count
- Only the highest model confidence
- Per-class errors and false-positive/false-negative counts
- When can a displayed 0.9 class score be treated as a calibrated 90% success estimate?
- Whenever a softmax layer exists
- Only after held-out calibration evidence supports that interpretation
- Whenever image resolution is high
- Whenever the network has many parameters
- What helps with an input that fits none of the trained classes?
- Always choose a known class with full certainty
- Add a new display label without training or validation
- An abstain or human-review route based on tested criteria
- Assume every unfamiliar photo is a camera failure