Complex-Valued Relational Embeddings With Amplitude-Phase Decomposition for Scene Graph Generation.
Abstract
Learning compact and discriminative representations of object relations is central to structured visual reasoning. In scene graph generation, these relations must be inferred from ambiguous visual cues under a heavy class imbalance; yet most methods encode a relation as a single real-valued vector that entangles interaction strength with a discriminative orientation, so frequent predicates overwrite the distinctions among rare ones. We propose CReaL, which fuses the subject-object pair with its context through complex multiplication, letting the context scale the magnitude and shift the phase of the relation. The result is split into an amplitude branch trained for interaction strength and a phase branch for the angular structure that discriminates similar relations, with branch-specific objectives that keep the two disentangled. A dynamic gating mechanism then fuses them with per-class weights computed from both branches, and the framework stays compatible with standard real-valued pipelines. On Visual Genome 150 (VG-150), CReaL improves mR@100 across predicate classification (PredCls), scene graph classification (SGCls), and scene graph detection (SGDet) over state-of-the-art methods, including a 5.9% gain in PredCls over the strongest knowledge-graph-based baseline despite using no external knowledge graph, and it attains the best overall score on Open Images v6 (OI-V6). Visualization and clustering analyses show that the amplitude and phase form complementary structures whose gated combination sharpens decision boundaries, indicating that explicit relational decomposition is an effective foundation for fine-grained relational understanding.