Vision-Language Grounding as Bidirectional Concept Correspondence
ConCor-1 treats grounding as finding every text-span and image-segment match, without being handed the phrase first.
The paper frames vision-language grounding as bidirectional concept correspondence across an image and its paired text. Its model uses learnable bridge tokens to predict a text mask, an image mask, and whether a correspondence exists. The authors convert grounding and segmentation datasets into one correspondence format for training and evaluation. They report ConCor-1 beating baselines by 48% correspondence F1 on a long-caption dataset and 29% on zero-shot LVIS. HF Daily Papers' note
The paper frames vision-language grounding as bidirectional concept correspondence across an image and its paired text. Its model uses learnable bridge tokens to predict a text mask, an image mask, and whether a correspondence exists. The authors convert grounding and segmentation datasets into one correspondence format for training and evaluation. They report ConCor-1 beating baselines by 48% correspondence F1 on a long-caption dataset and 29% on zero-shot LVIS. HF Daily Papers' note
score 5