Open Vocabulary Semantic Segmentation with Patch Aligned Contrastive Learning

TL;DR

This paper presents Patch Aligned Contrastive Learning (PACL), a novel method that aligns vision patch tokens with text tokens in CLIP, enabling open vocabulary semantic segmentation and improving zero-shot classification without additional annotations.

Contribution

PACL introduces a new compatibility function for CLIP that aligns patch and text tokens, enabling open vocabulary segmentation and enhancing zero-shot classification performance.

Findings

Achieves state-of-the-art zero-shot segmentation on four benchmarks.

Improves zero-shot classification accuracy across 12 datasets.

Enables segmentation without segmentation annotations during training.

Abstract

We introduce Patch Aligned Contrastive Learning (PACL), a modified compatibility function for CLIP's contrastive loss, intending to train an alignment between the patch tokens of the vision encoder and the CLS token of the text encoder. With such an alignment, a model can identify regions of an image corresponding to a given text input, and therefore transfer seamlessly to the task of open vocabulary semantic segmentation without requiring any segmentation annotations during training. Using pre-trained CLIP encoders with PACL, we are able to set the state-of-the-art on the task of open vocabulary zero-shot segmentation on 4 different segmentation benchmarks: Pascal VOC, Pascal Context, COCO Stuff and ADE20K. Furthermore, we show that PACL is also applicable to image-level predictions and when used with a CLIP backbone, provides a general improvement in zero-shot classification accuracy compared to CLIP, across a suite of 12 image classification datasets.