Universal Scene Graph Generation
This work addresses the limitation of single-modality scene graph generation for applications in multimodal AI, offering a more holistic representation but is incremental as it builds on existing scene graph research.
The paper tackles the problem of generating scene graphs from multiple modalities by introducing Universal Scene Graph (USG) and a parser, USG-Par, which addresses cross-modal alignment and domain imbalances, resulting in improved scene semantics expression and higher performance compared to single-modality methods.
Scene graph (SG) representations can neatly and efficiently describe scene semantics, which has driven sustained intensive research in SG generation. In the real world, multiple modalities often coexist, with different types, such as images, text, video, and 3D data, expressing distinct characteristics. Unfortunately, current SG research is largely confined to single-modality scene modeling, preventing the full utilization of the complementary strengths of different modality SG representations in depicting holistic scene semantics. To this end, we introduce Universal SG (USG), a novel representation capable of fully characterizing comprehensive semantic scenes from any given combination of modality inputs, encompassing modality-invariant and modality-specific scenes. Further, we tailor a niche-targeting USG parser, USG-Par, which effectively addresses two key bottlenecks of cross-modal object alignment and out-of-domain challenges. We design the USG-Par with modular architecture for end-to-end USG generation, in which we devise an object associator to relieve the modality gap for cross-modal object alignment. Further, we propose a text-centric scene contrasting learning mechanism to mitigate domain imbalances by aligning multimodal objects and relations with textual SGs. Through extensive experiments, we demonstrate that USG offers a stronger capability for expressing scene semantics than standalone SGs, and also that our USG-Par achieves higher efficacy and performance.