引導式測驗 · 中等 等級

AudioSet 與弱標記音訊事件測驗

測試 AudioSet 十秒多標籤剪輯、弱時間標籤、進化本體、消失源和特定任務的評估。

相關引導路徑Audioset 資料集
問題 1 的 8

Why can one AudioSet clip have more than one label?

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. Why can one AudioSet clip have more than one label?
    • The ontology requires every clip to contain exactly one event
    • A label is assigned to a particular video rather than an audio excerpt
    • Overlapping sounds are removed before annotation
    • Several sound events may occur in the same ten seconds
  2. What does a clip-level “dog bark” tag not provide by itself?
    • Evidence that a bark occurs somewhere in the segment
    • The exact onset and offset of every bark
    • A class ID associated with the clip
    • A label a tagging model can learn from
  3. Why record AudioSet ontology and release versions?
    • Every version necessarily has identical classes
    • Versioning proves each label is flawless
    • Class definitions and available metadata can change
    • Only the image thumbnail depends on release
  4. What can happen when a team rebuilds the audio from YouTube IDs later?
    • Some source clips may be unavailable, changing the usable subset
    • Unavailable videos should be counted as silent negatives
    • Every originally labeled video remains downloadable indefinitely
    • A missing clip can be reconstructed exactly from its ontology ID
  5. Which task requires more than AudioSet clip-level tags for direct evaluation?
    • Broad clip audio tagging
    • Multi-label presence prediction
    • Ranking likely event categories
    • Precise event-boundary localization
  6. Why should an unavailable source clip not be treated as a negative example?
    • The original human label becomes automatically false
    • A missing clip is equivalent to a silent recording
    • Missing media provide no evidence that the event was absent
    • The ontology automatically relabels unavailable IDs as negatives
  7. What should a product test beyond a web-clip benchmark?
    • Audio from its actual microphones and environments
    • Only the number of ontology classes
    • Only the size of CSV files
    • Only video titles
  8. Which claim is unsupported by high AudioSet tagging accuracy alone?
    • That it can output clip-level event scores
    • That the system can isolate exact source waveforms or timestamps
    • That a hierarchical ontology was used
    • That some held-out clips were tagged correctly