<ul><li>Researchers introduce SCENIR, a novel unsupervised scene graph-based retrieval framework emphasizing semantic content over low-level visual features.</li><li>SCENIR utilizes a Graph Autoencoder-based approach to eliminate the need for labeled training data, achieving superior performance and runtime efficiency compared to existing models.</li><li>The framework leverages Graph Edit Distance (GED) as a more reliable measure for scene graph similarity, replacing inconsistent caption-based supervision in image-to-image retrieval evaluation.</li><li>SCENIR demonstrates generalizability by applying it to unannotated datasets through automated scene graph generation and contributes to advancing state-of-the-art in counterfactual image retrieval.</li></ul>

SCENIR: Visual Semantic Clarity through Unsupervised Scene Graph Retrieval

Discover more