Disentangled Motif-aware Graph Learning for Phrase Grounding

In this paper, we propose a novel graph learning framework for phrase grounding in the image. Developing from sequential to dense model, existing works capture coarse-grained context but fail distinguish diversity of among phrases and image regions. contrast, pay special attention different motifs implied scene devise disentangled network integrate motif-aware contextual information into representations. Besides, adopt interventional strategies at feature structure levels consolidate generalize Finally, cross-modal is utilized fuse intra-modal features, where each can be computed similarity with regions select best-grounded one. We validate efficiency (DIGN) through series ablation studies, our model achieves state-of-the-art performance on Flickr30K Entities ReferIt Game benchmarks.