Title: QPredSGG: Leveraging Hybrid Quantum Predicate Learning for Long-Tailed Scene Graph Generation

Abstract:

Scene Graph Generation (SGG) hinges on the ability to perform relational reasoning regarding objects and their interactions; however, its efficacy is frequently hampered by significant long-tail predicate imbalance. Traditional SGG models tend to depend heavily on dataset statistics, which results in predictions that skew toward common relations rather than capturing nuanced semantic predicates. While current debiasing techniques have enhanced mean recall, predicate classification within these frameworks typically relies on substantial classical decision modules, incurring high parameter costs.

This study presents a hybrid quantum predicate classifier designed for SGG. We substitute the classical predicate head in the Causal Feature Enhancement Network (CFEN) with a Quantum Predicate Head (QP-Head), which is optimized via weighted cross-entropy. To the best of our knowledge, this represents one of the initial investigations into assessing a hybrid quantum architecture for predicate classification on the Visual Genome 150 dataset. We analyze how various factors—such as qubit count, encoding methods, entangling structures, and circuit depth—influence relational prediction outcomes.

The top-performing configuration, a 4-qubit QP-Head, employs Amplitude Embedding and Strongly Entangling Layers to distill 4096-dimensional pair features into a 16-dimensional quantum-compatible format, achieving a 256-fold reduction in dimensionality. This approach yields a mean Recall (mR@100) of 57.25%, a substantial improvement over the 41.1% achieved by the classical CFEN baseline, all while utilizing merely 96 trainable quantum parameters. When scaled to 8 qubits, the model sustains robust long-tail performance, attaining an mR@100 of 55.38% with 384 quantum parameters. Furthermore, our depth analysis highlights a balance between model expressibility and computational runtime. These findings indicate that compact hybrid quantum predicate heads can facilitate parameter-efficient, long-tail relational classification in intricate visual reasoning applications.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC