WeakMCN: Multi-task Collaborative Network for Weakly Supervised Referring Expression Comprehension and Segmentation
2026-09-07 12:00Models🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
4mentions
SummaryAI generated
The research team proposed the WeakMCN multi-task collaboration network, aimed at improving both the performance in weak supervision representation understanding (WREC) and segmentation (WRES). This network adopts a two-branch architecture, where the WREC branch is based on contrastive learning with anchors and supervises the WRES branch. Innovatively, dynamic visual feature enhancement (DVFE) and collaborative consistency module (CCM) were introduced. Experiments showed that on the RefCOCO, RefCOCO+, and RefCOCOg benchmarks, the performance in WREC and WRES increased by 3.91% and 13.11%, respectively. Additionally, under semi-supervised settings, on RefCOCO with only 1% labeled data, the performance in semi-symbolic understanding and segmentation increased by 8.94% and 7.71%, respectively, verifying its strong generalization ability in weak supervision scenarios and its superior performance compared to existing single-task methods.