GeneCIS: A Benchmark for General Conditional Image Similarity

TL;DR

GeneCIS introduces a benchmark for evaluating models' ability to adapt to various notions of image similarity in a zero-shot setting, revealing limitations of current models and proposing a scalable improvement method.

Contribution

The paper presents GeneCIS, a new benchmark for dynamic similarity notions, and a scalable method to enhance zero-shot image similarity performance.

Findings

Powerful CLIP models perform poorly on GeneCIS.

Performance on GeneCIS is weakly correlated with ImageNet accuracy.

The proposed method improves zero-shot similarity and surpasses supervised models on MIT-States.

Abstract

We argue that there are many notions of 'similarity' and that models, like humans, should be able to adapt to these dynamically. This contrasts with most representation learning methods, supervised or self-supervised, which learn a fixed embedding function and hence implicitly assume a single notion of similarity. For instance, models trained on ImageNet are biased towards object categories, while a user might prefer the model to focus on colors, textures or specific elements in the scene. In this paper, we propose the GeneCIS ('genesis') benchmark, which measures models' ability to adapt to a range of similarity conditions. Extending prior work, our benchmark is designed for zero-shot evaluation only, and hence considers an open-set of similarity conditions. We find that baselines from powerful CLIP models struggle on GeneCIS and that performance on the benchmark is only weakly correlated with ImageNet accuracy, suggesting that simply scaling existing methods is not fruitful. We further propose a simple, scalable solution based on automatically mining information from existing image-caption datasets. We find our method offers a substantial boost over the baselines on GeneCIS, and further improves zero-shot performance on related image retrieval benchmarks. In fact, though evaluated zero-shot, our model surpasses state-of-the-art supervised models on MIT-States. Project page at https://sgvaze.github.io/genecis/.