实验室2篇论文被2026 ACM MM接收。该会议是CCF A会议之一，每年举办一届。该届会议将于2026年11月10-14日于巴西里约热内卢举行。

题目: O2MSNet: One-to-Many Image-Text Alignment with Keyword Reconstruction for Multi-modal Medical Image Segmentation

作者：Qi Wen, Chengjie Luo, Jiezhou He, Zhiming Luo, Shaozi Li

摘要：Diagnostic reports provide rich semantic cues that can be leveraged to substantially enhance medical image segmentation. Existing methods typically adopt a one-to-one image-text matching paradigm, while in clinical practice, multiple images may share similar or even identical reports, forming an inherent one-to-many relationship, which limits the ability to capture shared semantics and leads to suboptimal lesion localization. To this end, we propose a novel multi-modal medical image segmentation framework (O2MSNet) that explicitly models the one-to-many correspondence between images and diagnostic reports. Specifically, we propose a text-guided IoU-weighted one-to-many cross-modal contrastive learning strategy (O2M-CCL) to associate images sharing the same report at the image level while enforcing keyword-guided pixel-level alignment, thereby overcoming the limitation of one-to-one matching and improving fine-grained semantic consistency. We further introduce a Keyword Reconstruction Constraint (KRC) to reconstruct masked diagnostic entities, enabling explicit modeling of key semantics (e.g., lesion quantity and location) and alleviating the lack of entity-level supervision. Moreover, we design a Progressive Hierarchical Cross-modal Interaction module to progressively fuse multi-scale visual features with textual semantics, enhancing cross-modal interaction and representation learning. Extensive experiments on QaTa-COV19 and MosMedData+ demonstrate that our method outperforms state-of-the-art approaches. All code will be released upon acceptance.

题目: From Frame to Sequence: Sequence-based Cross-view UAV Geo-localization via Robust Temporal Aggregation

作者：Jinliang Lin, Zhou Weikun, Siyi Sun, Hao Ruan, Zhuofan Xie, Zhiming Luo, Shaozi Li, Yu Zang, Cheng Wang

摘要：Diagnostic reports provide rich semantic cues that can be leveraged to substantially enhance medical image segmentation. Existing methods typically adopt a one-to-one image-text matching paradigm, while in clinical practice, multiple images may share similar or even identical reports, forming an inherent one-to-many relationship, which limits the ability to capture shared semantics and leads to suboptimal lesion localization. To this end, we propose a novel multi-modal medical image segmentation framework (O2MSNet) that explicitly models the one-to-many correspondence between images and diagnostic reports. Specifically, we propose a text-guided IoU-weighted one-to-many cross-modal contrastive learning strategy (O2M-CCL) to associate images sharing the same report at the image level while enforcing keyword-guided pixel-level alignment, thereby overcoming the limitation of one-to-one matching and improving fine-grained semantic consistency. We further introduce a Keyword Reconstruction Constraint (KRC) to reconstruct masked diagnostic entities, enabling explicit modeling of key semantics (e.g., lesion quantity and location) and alleviating the lack of entity-level supervision. Moreover, we design a Progressive Hierarchical Cross-modal Interaction module to progressively fuse multi-scale visual features with textual semantics, enhancing cross-modal interaction and representation learning. Extensive experiments on QaTa-COV19 and MosMedData+ demonstrate that our method outperforms state-of-the-art approaches. All code will be released upon acceptance.