{"type":"ScholarlyArticle","title":"Object-Centric Alignment and Anchor Distillation for Weakly Supervised Referring Expression Comprehension","authors":[{"name":"Yi Tian","slug":"yi-tian","given":"Yi","family":"Tian","affiliation":""},{"name":"Cheng Yang","slug":"cheng-yang","given":"Cheng","family":"Yang","affiliation":""},{"name":"Qingbao Huang","slug":"qingbao-huang","given":"Qingbao","family":"Huang","affiliation":""}],"abstract":"Weakly supervised Referring Expression Comprehension (WREC) aims to localize referred objects\r\nusing only image-text pairs without box-level annotations. Existing one-stage methods predominantly\r\nrely on anchor-level alignment, which suffers from\r\ntwo fundamental limitations: (1) anchors represent\r\nlocal visual patches rather than holistic objects, and\r\n(2) they lack the capability to model inter-object\r\nrelations. To address these issues, we propose\r\nOCAAD, an Object-Centric Alignment and Anchor Distillation framework. Our key insight is that\r\ndifferent self-attention heads in DINOv2 naturally\r\nattend to distinct semantic regions, effectively capturing object-level information. Building on this,\r\nOCAAD introduces two complementary modules:\r\n(1) Anchor-Object Distillation Module (AODM),\r\nwhich bridges the semantic gap between anchors\r\nand objects by transferring object-level knowledge\r\nfrom attention heads to anchors via overlap-aware\r\ncontrastive learning; and (2) Intra-Modal Relation Consistency (IRC), which explicitly models\r\ninter-object relations by enforcing the relational\r\nstructure among linguistic entities to match that of\r\ntheir visual counterparts. Extensive experiments on\r\nRefCOCO, RefCOCO+, and RefCOCOg demonstrate that OCAAD achieves new state-of-the-art\r\nperformance, validating the effectiveness of objectcentric alignment for WREC. Code is available at\r\nhttps://github.com/VILAN-Lab/OCAAD.","doi":"10.24963/ijcai.2026/190","url":"https://ijcai.org/proceedings/2026/190","pdf":"https://ijcai.org/proceedings/2026/0190.pdf","markdown":"https://ijcai.org/proceedings/2026/190.md","bibtex":"https://ijcai.org/proceedings/2026/bibtex/190","volume":{"slug":"2026","title":"Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence","short_title":"IJCAI 2026","url":"https://ijcai.org/proceedings/2026/","year":2026,"location":"Bremen, Germany","dates":"15-21 August 2026","editors":"Diego Calvanese","isbn":"978-1-956792-09-6","publisher":"International Joint Conferences on Artificial Intelligence Organization","doi_prefix":"10.24963/ijcai.2026"},"section":"Computer Vision","track":"Main Track","number":190,"pages":{"first":1704,"last":1712},"keywords":["Machine Learning: Weakly supervised learning","Computer Vision: Vision, language and reasoning"],"videos":[],"references":[],"license_note":"Copyright International Joint Conferences on Artificial Intelligence."}