{"type":"ScholarlyArticle","title":"PatternCIR Benchmark and TisCIR: Advancing Zero-Shot Composed Image Retrieval in Remote Sensing","authors":[{"name":"Zhechun Liang","slug":"zhechun-liang","given":"Zhechun","family":"Liang","affiliation":""},{"name":"Tao Huang","slug":"tao-huang","given":"Tao","family":"Huang","affiliation":""},{"name":"Fangfang Wu","slug":"fangfang-wu","given":"Fangfang","family":"Wu","affiliation":""},{"name":"Shiwen Xue","slug":"shiwen-xue","given":"Shiwen","family":"Xue","affiliation":""},{"name":"Zhenyu Wang","slug":"zhenyu-wang","given":"Zhenyu","family":"Wang","affiliation":""},{"name":"Weisheng Dong","slug":"weisheng-dong","given":"Weisheng","family":"Dong","affiliation":""},{"name":"Xin Li","slug":"xin-li","given":"Xin","family":"Li","affiliation":""},{"name":"Guangming Shi","slug":"guangming-shi","given":"Guangming","family":"Shi","affiliation":""}],"abstract":"Remote sensing composed image retrieval\r\n(RSCIR) is a new vision-language task that takes\r\na composed query of an image and text, aiming to\r\nsearch for a target remote sensing image satisfying\r\ntwo conditions from intricate remote sensing\r\nimagery. However, the existing attribute-based\r\nbenchmark Patterncom in RSCIR has significant\r\nflaws, including the lack of query text sentences\r\nand paired triplets, thus making it unable to evalu\u0002ate the latest methods. To address this, we propose\r\nthe Zero-Shot Query Text Generator (ZS-QTG)\r\nthat can generate full query text sentences based on\r\nattributes, and then, by capitalizing on ZS-QTG,\r\nwe develop the PatternCIR benchmark. Pattern\u0002CIR rectifies Patterncom’s deficiencies and enables\r\nthe evaluation of existing methods. Additionally,\r\nwe explore zero-shot composed image retrieval\r\nmethods that do not rely on massive pre-collected\r\ntriplets for training. Existing methods use only\r\nthe text during retrieval, performing poorly in\r\nRSCIR. To improve this, we propose Text-image\r\nSequential Training of Composed Image Retrieval\r\n(TisCIR). TisCIR undergoes sequential training of\r\nmultiple self-masking projection and fine-grained\r\nimage attention modules, which endows it with\r\nthe capacity to filter out conflicting information\r\nbetween the image and text, enhancing the retrieval\r\nby utilizing both modalities in harmony. TisCIR\r\noutperforms existing methods by 12.40% to\r\n62.03% on PatternCIR, achieving state-of-the-art\r\nperformance in RSCIR. The data and code are\r\navailable here.","doi":"10.24963/ijcai.2025/171","url":"https://ijcai.org/proceedings/2025/171","pdf":"https://ijcai.org/proceedings/2025/0171.pdf","markdown":"https://ijcai.org/proceedings/2025/171.md","bibtex":"https://ijcai.org/proceedings/2025/bibtex/171","volume":{"slug":"2025","title":"Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence","short_title":"IJCAI 2025","url":"https://ijcai.org/proceedings/2025/","year":2025,"location":"Montreal, Canada","dates":"16-22 August 2025","editors":"James Kwok","isbn":"978-1-956792-06-5","publisher":"International Joint Conferences on Artificial Intelligence Organization","doi_prefix":"10.24963/ijcai.2025"},"section":"Computer Vision","track":"Main Track","number":171,"pages":{"first":1530,"last":1538},"keywords":["Computer Vision: CV: Scene analysis and understanding","Computer Vision: CV: Image and video retrieval","Computer Vision: CV: Multimodal learning","Computer Vision: CV: Vision, language and reasoning"],"videos":[],"references":[],"license_note":"Copyright International Joint Conferences on Artificial Intelligence."}