VLM-KD: Knowledge Distillation from VLM for Long-Tail Visual Recognition

TL;DR

This paper presents VLM-KD, a novel knowledge distillation method that leverages off-the-shelf vision-language models to generate text supervision, significantly improving long-tail visual recognition performance.

Contribution

It introduces a framework for distilling free-form text from VLMs into vision encoders, a first in applying text supervision from off-the-shelf VLMs to vanilla vision models.

Findings

Outperforms state-of-the-art long-tail classifiers on benchmark datasets.

Demonstrates effectiveness of text supervision from VLMs in knowledge distillation.

First to utilize off-the-shelf VLMs for distillation into randomly initialized vision encoders.

Abstract

For visual recognition, knowledge distillation typically involves transferring knowledge from a large, well-trained teacher model to a smaller student model. In this paper, we introduce an effective method to distill knowledge from an off-the-shelf vision-language model (VLM), demonstrating that it provides novel supervision in addition to those from a conventional vision-only teacher model. Our key technical contribution is the development of a framework that generates novel text supervision and distills free-form text into a vision encoder. We showcase the effectiveness of our approach, termed VLM-KD, across various benchmark datasets, showing that it surpasses several state-of-the-art long-tail visual classifiers. To our knowledge, this work is the first to utilize knowledge distillation with text supervision generated by an off-the-shelf VLM and apply it to vanilla randomly initialized vision encoders.