Hallucination In Object Detection -- A Study In Visual Part Verification

TL;DR

This paper investigates how object detectors can hallucinate objects and parts, leading to false positives in visual part verification, and introduces a new dataset for evaluating this issue.

Contribution

It reveals hallucination issues in object detectors for part verification and presents the first dataset, DelftBikes, for systematic evaluation of this problem.

Findings

Object detectors can hallucinate missing objects and parts.

DelftBikes dataset provides detailed annotations for parts and their presence.

Evaluation shows variability in detector performance on part verification.

Abstract

We show that object detectors can hallucinate and detect missing objects; potentially even accurately localized at their expected, but non-existing, position. This is particularly problematic for applications that rely on visual part verification: detecting if an object part is present or absent. We show how popular object detectors hallucinate objects in a visual part verification task and introduce the first visual part verification dataset: DelftBikes, which has 10,000 bike photographs, with 22 densely annotated parts per image, where some parts may be missing. We explicitly annotated an extra object state label for each part to reflect if a part is missing or intact. We propose to evaluate visual part verification by relying on recall and compare popular object detectors on DelftBikes.