引導式測驗 · 中等 等級
弱監督物件定位測驗
檢查影像級監督、類別激活圖、判別性補丁、保留框、定位指標和精確度限制。
問題 1 的 8
What training annotation is commonly available in weakly supervised object localization?
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。
- What training annotation is commonly available in weakly supervised object localization?
- A precise pixel mask for each image
- A complete box for every target
- An image-level class label without a box for every object
- A verified 3D pose for each object
- Why can a CAM highlight only a bird’s head instead of the full bird?
- Image labels enforce a complete outline
- The camera cannot capture wings
- Global average pooling supplies human-drawn masks
- The head may be sufficient as a discriminative cue for the class score
- Which evidence is needed to assess localization quality?
- Only image-level class accuracy
- Held-out boxes or regions annotated independently
- Only the training image count
- Only the heat map color scheme
- A classifier has high accuracy but its maps highlight backgrounds. What follows?
- Class prediction and localization quality must be evaluated separately
- High class accuracy verifies boundaries
- Context highlighting cannot affect a classification
- Every heat map is a causal proof
- Why does converting a heat map to a bounding box require care?
- Image-level labels already provide exact corners
- Box size is independent of threshold
- Threshold and spatial resolution affect the proposed extent
- Every highlighted pixel belongs to the object
- Which task may need stronger supervision than a rough weakly supervised highlight?
- Suggesting an editable crop
- Helping inspect a broad class cue
- Showing an approximate region of interest
- Masking every visible face before publication
- How does the original CAM setup obtain class-specific spatial emphasis?
- Copy an expert-drawn mask into each image
- Combine spatial feature maps with weights tied to the target class
- Compare only file checksums
- Use a depth value at each pixel
- Why hold location-labeled evaluation images out of weakly supervised training?
- They test whether rough regions transfer without direct location teaching
- Location labels must train the classifier
- No location annotations can exist in a test set
- Holding out location labels makes evaluation impossible