VTuckeR: Multimodal Tucker Fusion for Scene Graph Generation

Originated from object segmentation and word vector representations, Scene Graph Generation (SGG) became a complex task built on enumerous research results. Today's scene graph generation (SGG) task is still far from practical, we believe that it's due to unfair matching strategies between images and object regions. Inspired by Tucker decomposition's success in VQA area, in this paper, we propose VTuckeR, a relatively straightforward but powerful linear model based on Tucker decomposition of the tensor representation of images and object features. In VTuckeR, we control the complexity of the merging scheme while keeping itself good interpretability. We show our model is able to enforce multiple types of Scene Graph Generation models in Visual Genome dataset in the PredCls mode. What's more, a more accurate scene graph may aid prediction of wireless channel dynamics in the future, which is called V2C.