Deciphering the Role of Representation Disentanglement: Investigating Compositional Generalization in CLIP Models

Overview

Welcome to the official repository for our ECCV 2024 accepted paper, "Deciphering the Role of Representation Disentanglement: Investigating Compositional Generalization in CLIP Models." This project explores the compositional out-of-distribution (C-OoD) generalization capabilities of CLIP models, focusing on how these models handle unseen combinations of known concepts.

Dataset ImageNet-AO

You can access the filtered version of our dataset (v2), processed using GPT-4-V, at the following link:

ImageNet-AO v2 (Filtered by GPT-4-V)