Ref-Diff: Zero-shot Referring Image Segmentation with Generative Models

Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, Wangmeng Zuo

Keywords: Zero-shot Referring Image Segmentation, Generative Model

TL;DR: This paper shows using a sole generative model can beats SOTA weakly-supervised models without a proposal generator, and combining both generative and discriminative models can surpasses competing methods by a substantial margin in zero-shot RIS.

Abstract: Zero-shot referring image segmentation (RIS) presents a significant challenge. It requires identifying an instance segmentation mask using referring descriptions, without having been trained on such paired data. While existing zero-shot RIS methods mainly utilize pre-trained discriminative models (e.g., CLIP), this study observes that generative models (e.g., Stable Diffusion) can discern relationships between various visual elements and text descriptions, an area yet to be explored in this task. In this work, we introduce the Referring Diffusional Segmentor (Ref-Diff), a model that leverages the fine-grained multi-modal information derived from generative models. Our findings show that even without an external proposal generator, our Ref-Diff with a sole generative model outperforms SOTA weakly-supervised models on RefCOCO+ and RefCOCOg. Notably, when combining both generative and discriminative models, our Ref-Diff+ surpasses competing methods by a substantial margin. This highlights the constructive role of generative models in this domain, providing complementary advantages alongside discriminative models to enhance referring segmentation. Our source code will be publicly available.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4572