Anticipating Future Object Compositions without Forgetting

TL;DR

This paper advances compositional zero-shot learning in object detection by integrating grounding, soft prompting, anticipation, and contrastive tuning, significantly improving generalization to novel object-attribute combinations.

Contribution

It introduces a novel framework combining compositional soft prompting, anticipation, and contrastive tuning to enhance object detection in CZSL without forgetting prior knowledge.

Findings

70.5% improvement over CSP on harmonic mean in CLEVR

14.5% increase in harmonic mean across datasets

Effective learning of compositions with limited data

Abstract

Despite the significant advancements in computer vision models, their ability to generalize to novel object-attribute compositions remains limited. Existing methods for Compositional Zero-Shot Learning (CZSL) mainly focus on image classification. This paper aims to enhance CZSL in object detection without forgetting prior learned knowledge. We use Grounding DINO and incorporate Compositional Soft Prompting (CSP) into it and extend it with Compositional Anticipation. We achieve a 70.5% improvement over CSP on the harmonic mean (HM) between seen and unseen compositions on the CLEVR dataset. Furthermore, we introduce Contrastive Prompt Tuning to incrementally address model confusion between similar compositions. We demonstrate the effectiveness of this method and achieve an increase of 14.5% in HM across the pretrain, increment, and unseen sets. Collectively, these methods provide a framework for learning various compositions with limited data, as well as improving the performance of underperforming compositions when additional data becomes available.