Hongyu Li, Tianrui Hui, Zihan Ding, Jing Zhang, Bin Ma, Xiaoming Wei, Jizhong Han, Si Liu
{"title":"Dynamic Prompting of Frozen Text-to-Image Diffusion Models for Panoptic Narrative Grounding","authors":"Hongyu Li, Tianrui Hui, Zihan Ding, Jing Zhang, Bin Ma, Xiaoming Wei, Jizhong Han, Si Liu","doi":"arxiv-2409.08251","DOIUrl":null,"url":null,"abstract":"Panoptic narrative grounding (PNG), whose core target is fine-grained\nimage-text alignment, requires a panoptic segmentation of referred objects\ngiven a narrative caption. Previous discriminative methods achieve only weak or\ncoarse-grained alignment by panoptic segmentation pretraining or CLIP model\nadaptation. Given the recent progress of text-to-image Diffusion models,\nseveral works have shown their capability to achieve fine-grained image-text\nalignment through cross-attention maps and improved general segmentation\nperformance. However, the direct use of phrase features as static prompts to\napply frozen Diffusion models to the PNG task still suffers from a large task\ngap and insufficient vision-language interaction, yielding inferior\nperformance. Therefore, we propose an Extractive-Injective Phrase Adapter\n(EIPA) bypass within the Diffusion UNet to dynamically update phrase prompts\nwith image features and inject the multimodal cues back, which leverages the\nfine-grained image-text alignment capability of Diffusion models more\nsufficiently. In addition, we also design a Multi-Level Mutual Aggregation\n(MLMA) module to reciprocally fuse multi-level image and phrase features for\nsegmentation refinement. Extensive experiments on the PNG benchmark show that\nour method achieves new state-of-the-art performance.","PeriodicalId":501130,"journal":{"name":"arXiv - CS - Computer Vision and Pattern Recognition","volume":null,"pages":null},"PeriodicalIF":0.0000,"publicationDate":"2024-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"arXiv - CS - Computer Vision and Pattern Recognition","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/arxiv-2409.08251","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0
Abstract
Panoptic narrative grounding (PNG), whose core target is fine-grained
image-text alignment, requires a panoptic segmentation of referred objects
given a narrative caption. Previous discriminative methods achieve only weak or
coarse-grained alignment by panoptic segmentation pretraining or CLIP model
adaptation. Given the recent progress of text-to-image Diffusion models,
several works have shown their capability to achieve fine-grained image-text
alignment through cross-attention maps and improved general segmentation
performance. However, the direct use of phrase features as static prompts to
apply frozen Diffusion models to the PNG task still suffers from a large task
gap and insufficient vision-language interaction, yielding inferior
performance. Therefore, we propose an Extractive-Injective Phrase Adapter
(EIPA) bypass within the Diffusion UNet to dynamically update phrase prompts
with image features and inject the multimodal cues back, which leverages the
fine-grained image-text alignment capability of Diffusion models more
sufficiently. In addition, we also design a Multi-Level Mutual Aggregation
(MLMA) module to reciprocally fuse multi-level image and phrase features for
segmentation refinement. Extensive experiments on the PNG benchmark show that
our method achieves new state-of-the-art performance.