Hallucination In Object Detection

Overview

This work demonstrates that object detectors can incorrectly identify non-existent objects with precise localization. We address this through visual part verification - determining whether object components are present or absent.

Key Contributions

• Introduces DelftBikes: 10,000 bicycle photographs with 22 densely annotated parts per image
• First dataset with explicit labels for missing vs. intact object parts
• Systematic evaluation of popular object detectors on part verification
• Demonstrates detector hallucination of non-existent parts

Links

• arXiv
• PDF