InfoCLIP: Bridging Vision-Language Pretraining and Open-Vocabulary Semantic Segmentation via Information-Theoretic Alignment Transfer

Benchmark Model Rank Results
open-vocabulary-semantic-segmentation-on-ade20k-150InfoCLIP (CLIP ViT-B/16)–mIoU: 32.1
open-vocabulary-semantic-segmentation-on-ade20k-150InfoCLIP (CLIP ViT-L/14)–mIoU: 38.5
open-vocabulary-semantic-segmentation-on-ade20k-847InfoCLIP (CLIP ViT-B/16)–mIoU: 12.6
open-vocabulary-semantic-segmentation-on-ade20k-847InfoCLIP (CLIP ViT-L/14)–mIoU: 16.6
open-vocabulary-semantic-segmentation-on-pascal-context-459InfoCLIP (CLIP ViT-B/16)–mIoU: 19.5
open-vocabulary-semantic-segmentation-on-pascal-context-459InfoCLIP (CLIP ViT-L/14)–mIoU: 24.6
open-vocabulary-semantic-segmentation-on-pascal-context-59InfoCLIP (CLIP ViT-B/16)–mIoU: 58.1
open-vocabulary-semantic-segmentation-on-pascal-context-59InfoCLIP (CLIP ViT-L/14)–mIoU: 63.5
open-vocabulary-semantic-segmentation-on-pascalvoc-20InfoCLIP (CLIP ViT-B/16)–mIoU: 95.5
open-vocabulary-semantic-segmentation-on-pascalvoc-20InfoCLIP (CLIP ViT-L/14)–mIoU: 97.5