CounterFactual Retrieval for Evidence-Aware Individual Animal Identification
CFR tests whether animal identification models rely on the animal or on background cues by comparing retrieval across full-image, foreground-only, and shortcut-only views.
The **Counterfactual Foreground Retrieval (CFR)** protocol evaluates the same identification task on full images, animal-only foregrounds, and shortcut-only views. The authors release around **0.5M masks across 34 datasets**, plus evaluation and mask-generation code.
For image retrieval systems, add evidence checks alongside accuracy: compare performance with the subject isolated and removed to see whether background cues carry identity signal.
The **Counterfactual Foreground Retrieval (CFR)** protocol evaluates the same identification task on full images, animal-only foregrounds, and shortcut-only views. The authors release around **0.5M masks across 34 datasets**, plus evaluation and mask-generation code. For image retrieval systems, add evidence checks alongside accuracy: compare performance with the subject isolated and removed to see whether background cues carry identity signal. The work concerns individual animal identification. The material does not report specific results for other retrieval domains or establish that the protocol transfers unchanged.
CFR makes shortcut use directly testable by comparing identification on full images, subject-only foregrounds, and shortcut-only views. It complements visual stress and corruption evaluations with a controlled way to check whether image retrieval depends on the subject or surrounding cues. The evidence is for individual animal identification; transfer to other retrieval tasks remains unestablished.