Open Vocabulary Semantic Segmentation with Patch Aligned Contrastive Learning

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.

PDF Abstract CVPR 2023 PDF CVPR 2023 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Open Vocabulary Semantic Segmentation ADE20K-150 PACL mIoU 31.4 # 9
Open Vocabulary Semantic Segmentation Cityscape-171 PACL mIoU 38.8 # 1
Open Vocabulary Semantic Segmentation PASCAL Context-59 PACL mIoU 50.1 # 11
Open Vocabulary Semantic Segmentation PascalVOC-20 PACL mIoU 72.3 # 13