GeoMVC Model Achieves Top Ranks in Multimodal Misogyny Detection Challenge

A team of researchers introduced GeoMVC (Geometric Interaction and Multi-View Consensus), an innovative system designed to identify misogyny in internet memes. At the CC-MMD Grand Challenge during ICMI 2026, it secured Rank 2 in the Malayalam partition with a Macro F1 score of 0.892 and Rank 3 in the Chinese partition. This system tackles the complexities of memes that feature a semantic clash between visual elements and text, where hateful messages are often implicit and culturally specific. GeoMVC employs a Geometric Interaction Layer to model cross-modal alignment using Hadamard products and cosine similarity of visual and textual embeddings, enhancing beyond static feature concatenation. To address distribution shifts from noisy OCR and code-mixed transliteration, a Multi-View Consensus approach aggregates predictions from various text views. The findings were published on arXiv (2607.22709v1) and showcased at the CC-MMD Grand Challenge.