quiz bu lëkkale ak tegtal · Diggante Niveau

Dataset Bias in Computer Vision Quiz

Test collection bias, background shortcuts, near-duplicate leakage, external validation, subgroup slices, and limits of average accuracy.

Yooni njiit yu jëm ciDataset Bias In Computer Vision
Laaj 1 de 8

A wildlife model recognizes an animal only when snow is visible. What shortcut may it have learned?

Xoolaatal laaj bu nekk8 laaj

Sooy def quiz bi, tontu yi ak leeral yi dañuy feeñ.

  1. A wildlife model recognizes an animal only when snow is visible. What shortcut may it have learned?
    • An invariant anatomical cue from the animal
    • A feature equally present across all habitats
    • The animal’s identity independent of scenery
    • Background correlated with the animal label
  2. Why can random splitting of neighboring video frames overstate accuracy?
    • Near duplicates can appear in both training and test sets
    • It guarantees all rare classes disappear
    • It always changes model architecture
    • It ensures an independent camera source
  3. What did the Torralba and Efros dataset-bias work help show?
    • Dataset names alone establish transfer
    • Background cues never influence recognition
    • Visual datasets can have recognizable source signatures and weak cross-dataset transfer
    • One dataset contains every camera condition
  4. A model is accurate overall but fails under dim light. Which report makes that visible?
    • Only a repeated aggregate accuracy
    • Error broken down by lighting conditions with sample counts
    • Only training loss from the same lighting
    • Only the model parameter count
  5. Why is an external-site test useful after a single-site benchmark?
    • It probes changes in camera, scene and collection conditions
    • It guarantees every possible source bias is gone
    • It identifies the exact cause of every failure
    • It makes label-rule checks unnecessary
  6. Which comparison helps test whether background rather than object drives a prediction?
    • Compare only adjacent crops from one image
    • Keep every object on its original background
    • Show the object across varied backgrounds and backgrounds without it
    • Change model confidence without changing examples
  7. A small object is unlabeled in many test images. What risk follows?
    • All model confidences must decrease
    • Correct detections may be counted as false positives
    • Only training accuracy is affected
    • The missing box is automatically drawn by evaluation
  8. Which conclusion follows from high accuracy on one source distribution?
    • It proves identical errors across regions
    • It rules out all background shortcuts
    • It establishes fairness without subgroup tests
    • It supports that tested setting, not a promise for every deployment population