The Elephant in the Room

TL;DR

This paper reveals that state-of-the-art object detectors are vulnerable to 'object transplanting', where replacing image regions with other objects causes non-local detection failures, highlighting limitations in current models.

Contribution

The paper introduces the concept of 'object transplanting' to demonstrate non-local failures in object detectors and analyzes the underlying causes of these vulnerabilities.

Findings

Object transplanting causes detection failures.

Small positional changes affect object identity detection.

Non-local impacts challenge current object detector robustness.

Abstract

We showcase a family of common failures of state-of-the art object detectors. These are obtained by replacing image sub-regions by another sub-image that contains a trained object. We call this "object transplanting". Modifying an image in this manner is shown to have a non-local impact on object detection. Slight changes in object position can affect its identity according to an object detector as well as that of other objects in the image. We provide some analysis and suggest possible reasons for the reported phenomena.