From ImageNet to Image Classification: Contextualizing Progress on Benchmarks

TL;DR

This paper investigates how the design choices and noise in the ImageNet dataset creation process introduce biases and misalignments, affecting model evaluation and highlighting the need for improved benchmarking methods.

Contribution

It provides an analysis of the impact of data collection biases in ImageNet and offers refined annotations to better align benchmarks with real-world tasks.

Findings

Biases in ImageNet affect model performance evaluation.

Noisy data collection leads to systematic dataset-model misalignment.

Refined annotations improve benchmark fidelity.

Abstract

Building rich machine learning datasets in a scalable manner often necessitates a crowd-sourced data collection pipeline. In this work, we use human studies to investigate the consequences of employing such a pipeline, focusing on the popular ImageNet dataset. We study how specific design choices in the ImageNet creation process impact the fidelity of the resulting dataset---including the introduction of biases that state-of-the-art models exploit. Our analysis pinpoints how a noisy data collection pipeline can lead to a systematic misalignment between the resulting benchmark and the real-world task it serves as a proxy for. Finally, our findings emphasize the need to augment our current model training and evaluation toolkit to take such misalignments into account. To facilitate further research, we release our refined ImageNet annotations at https://github.com/MadryLab/ImageNetMultiLabel.