ViCo: Word Embeddings from Visual Co-occurrences

TL;DR

This paper introduces a method to learn word embeddings from visual co-occurrences in images, enhancing the representation of visual concepts beyond text-only embeddings and improving performance on vision-language tasks.

Contribution

The paper presents a novel approach to derive word embeddings from visual data, complementing text-based embeddings and demonstrating their effectiveness in multiple vision-language applications.

Findings

Visual co-occurrence-based embeddings improve concept similarity representation.

Augmenting GloVe with visual embeddings enhances vision-language task performance.

Random embeddings perform similarly to learned embeddings on supervised tasks.

Abstract

We propose to learn word embeddings from visual co-occurrences. Two words co-occur visually if both words apply to the same image or image region. Specifically, we extract four types of visual co-occurrences between object and attribute words from large-scale, textually-annotated visual databases like VisualGenome and ImageNet. We then train a multi-task log-bilinear model that compactly encodes word "meanings" represented by each co-occurrence type into a single visual word-vector. Through unsupervised clustering, supervised partitioning, and a zero-shot-like generalization analysis we show that our word embeddings complement text-only embeddings like GloVe by better representing similarities and differences between visual concepts that are difficult to obtain from text corpora alone. We further evaluate our embeddings on five downstream applications, four of which are vision-language tasks. Augmenting GloVe with our embeddings yields gains on all tasks. We also find that random embeddings perform comparably to learned embeddings on all supervised vision-language tasks, contrary to conventional wisdom.