引導式測驗 · 中等 等級

臉部偵測演算法測驗

檢查人臉框與身分、級聯和學習偵測器、置信度閾值、定位錯誤和下游隱私影響。

相關引導路徑臉部檢測演算法
問題 1 的 8

A detector returns a rectangle around a face. Which fact does that output establish by itself?

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 detector returns a rectangle around a face. Which fact does that output establish by itself?
    • An image region is likely to contain a face
    • That the region matches an enrolled person
    • That the presenter is physically live
    • That the region shows a particular emotional state
  2. Why do cascade classifiers reject many windows in early stages?
    • To verify identity before checking the next stage
    • To reserve later classifiers for already rejected windows
    • To save computation on regions unlikely to contain a face
    • To ensure a fixed number of faces in each image
  3. Which design feature is documented for RetinaFace?
    • It uses only a fixed hand-coded cascade without landmarks
    • It uses face-box and landmark supervision in a learned detector
    • It verifies identity from the detected box by itself
    • It requires a known face at the image center
  4. If a team lowers its face-score threshold, what tradeoff should it expect?
    • Fewer candidate boxes will pass and misses may increase
    • Box coordinates become more precise without retraining
    • All surviving boxes are guaranteed to be real faces
    • More candidate faces may be found, with more false alarms possible
  5. Why does a missed detection matter especially in a face-blurring tool?
    • The blur becomes sharper around every detected face
    • A false positive can reveal an undetected person
    • The unlocated face may remain visible in the released image
    • An identity matcher automatically fills the missing box
  6. Two boxes overlap almost completely on one face. What postprocessing often handles this?
    • Suppressing redundant overlapping predictions
    • Changing the person’s enrolled identity template
    • Increasing the face-score threshold for all boxes
    • Treating every overlapping box as a separate person
  7. Which evaluation slice can reveal failures hidden by an overall average?
    • Only accuracy on large frontal faces
    • The average score over all image pixels
    • Performance only on training-session frames
    • Accuracy for tiny, occluded and side-view faces
  8. How should frames from one recording session be split for a realistic benchmark?
    • Put consecutive frames into different partitions
    • Keep related frames together to reduce near-duplicate leakage
    • Train on all frames and retest the same scenes
    • Select the threshold using labels from the final test