引導式測驗 · 中等 等級
ImageNet-C 與 Common-Corrupt 穩健性測驗
測試 ImageNet-C 損壞和嚴重性、乾淨與移位的準確性、每次損壞報告、過度擬合和部署限制。
問題 1 的 8
Why does ImageNet-C alter validation images while retaining their object labels?
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。
- Why does ImageNet-C alter validation images while retaining their object labels?
- To create new categories for altered images
- To test predictions under input corruptions without changing the intended category
- To make synthetic effects identical to all field weather
- To replace the original clean evaluation
- Which case belongs to the common-corruption question rather than a targeted adversarial attack?
- Optimizing each input to fool a chosen model
- Changing reference labels after inspecting outputs
- Updating model weights during the test
- Evaluating blur at a specified severity across many images
- A model’s clean accuracy increases. What cannot be concluded automatically?
- Its accuracy improved under every ImageNet-C corruption
- Corruption accuracy must improve equally
- Per-severity error must fall for all effects
- The model is calibrated on real deployment data
- Why report results by corruption type and severity?
- A mean already exposes every severe failure
- Severity slices change the underlying object label
- An average can hide a severe failure under one disturbance
- Detailed slices remove the need for field tests
- Why is synthetic snow on ImageNet-C not a full storm-camera test?
- Synthetic snow cannot affect an image model
- Real exposure, sensor processing and weather interactions may differ
- Synthetic weather changes the class ontology
- Real cameras always produce the same pixel pattern
- A team tunes repeatedly on public ImageNet-C labels. What evaluation risk grows?
- Overfitting the finite benchmark transformations
- The public test automatically becomes larger
- Repeated use guarantees out-of-domain robustness
- The model becomes independent of augmentation
- Which additional test supports a warehouse camera deployment?
- Only the public benchmark mean score
- Only clean consumer photos from the same source
- Only the architecture name and parameter count
- Images corrupted by the actual device and operating conditions
- What should a system do when an image is too degraded for a confident decision?
- Trust every low-confidence guess
- Assume a public benchmark pass covers the poor input
- Use a defined fallback such as recapture or abstention
- Treat blur as proof that the target class changed