MaskInversion: Localized Embeddings via Optimization of Explainability Maps

TL;DR

MaskInversion introduces a novel optimization-based method to generate localized, context-aware image region embeddings using explainability maps from pre-trained vision-language models, enhancing tasks like retrieval and captioning.

Contribution

It presents MaskInversion, a new approach that refines region embeddings by optimizing explainability maps, applicable to any pre-trained foundation model without retraining.

Findings

Effective in localized image understanding tasks

Outperforms state-of-the-art methods on multiple datasets

Enables broad applications like captioning and image generation

Abstract

Vision-language foundation models such as CLIP have achieved tremendous results in global vision-language alignment, but still show some limitations in creating representations for specific image regions. % To address this problem, we propose MaskInversion, a method that leverages the feature representations of pre-trained foundation models, such as CLIP, to generate a context-aware embedding for a query image region specified by a mask at test time. MaskInversion starts with initializing an embedding token and compares its explainability map, derived from the foundation model, to the query mask. The embedding token is then subsequently refined to approximate the query region by minimizing the discrepancy between its explainability map and the query mask. During this process, only the embedding vector is updated, while the underlying foundation model is kept frozen allowing to use MaskInversion with any pre-trained model. As deriving the explainability map involves computing its gradient, which can be expensive, we propose a gradient decomposition strategy that simplifies this computation. The learned region representation can be used for a broad range of tasks, including open-vocabulary class retrieval, referring expression comprehension, as well as for localized captioning and image generation. We evaluate the proposed method on all those tasks on several datasets such as PascalVOC, MSCOCO, RefCOCO, and OpenImagesV7 and show its capabilities compared to other SOTA approaches.