[Paper Review] The Elephant in the Room

Amir Rosenfeld, Richard S. Zemel+1 more|arXiv (Cornell University)|Jan 1, 2018

Adversarial Robustness in Machine Learning14 references78 citations

TL;DR

This paper investigates the vulnerability of state-of-the-art object detectors to object transplanting—replacing image regions with a different object from another image. It demonstrates that such non-local perturbations cause unstable detection, incorrect object identities, and non-local effects on unrelated objects, primarily due to feature interference from ROI pooling and lack of robust contextual reasoning.

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.

Motivation & Objective

• To investigate the robustness of state-of-the-art object detectors under non-local, non-perturbative image manipulations.
• To identify failure modes in object detection when objects are transplanted into new spatial contexts.
• To analyze the root causes of detection instability and misclassification under such perturbations.
• To evaluate the role of feature interference, non-maximum suppression, and contextual reasoning in detector failures.

Proposed method

• Transplanted objects from COCO validation images into new locations in other images using ground-truth segmentation masks.
• Generated test images by varying object positions in a grid-like fashion (step size k=10) to ensure full visibility.
• Compared detector outputs (detections with bounding boxes, scores, and classes) between original and transplanted images.
• Used the 'Class-Matching' criterion to rank images based on the number of new or missing detected classes.
• Evaluated multiple state-of-the-art models from the TensorFlow Object Detection API, including Faster R-CNN with NASNet and Mask R-CNN.
• Analyzed effects through qualitative and quantitative analysis, focusing on feature interference, NMS side effects, and contextual reasoning failures.

Experimental results

Research questions

• RQ1 How do object detectors respond to the transplantation of an object into a new image location?
• RQ2 Why do transplanted objects sometimes disappear, change identity, or cause other objects to vanish or switch class?
• RQ3 To what extent do non-local effects—such as changes in distant objects—arise from detector architecture components?
• RQ4 What role does feature interference from ROI pooling play in detector instability?
• RQ5 Can contextual reasoning failures explain the observed misclassifications in non-co-occurring object configurations?

Key findings

• Transplanted objects were frequently undetected or misclassified as different classes (e.g., an elephant detected as a chair) depending on position.
• Objects not overlapping with the transplanted object were sometimes undetected or changed identity, demonstrating non-local effects.
• The strongest detector (Faster R-CNN NAS COCO, mAP 43%) showed significant misclassification chains: e.g., a hot-dog became a sandwich, and a sign was detected as a book.
• Non-maximum suppression (NMS) contributed to chain reactions, where suppression of one object led to reactivation of others far from the transplant.
• Feature interference from ROI pooling—especially pooling from background and surrounding regions—was identified as a primary cause of instability.
• Even in same-image transplanting (duplicating an object), similar failures occurred, indicating that context and spatial layout heavily influence detection outcomes.