The paper addresses the lack of formal definitions for scene graph generation metrics in the literature. It provides comprehensive and formal definitions for commonly used metrics such as Recall@k, Mean Recall@k, Pair Recall@k, and No Graph Constraint Recall@k, accompanied by pseudo-code for better understanding.
The authors also introduce an efficient Python package called SGBench that implements all the defined metrics in a lightweight and easy-to-use manner. SGBench is designed to be more efficient and have less boilerplate code compared to existing implementations.
Furthermore, the authors present a public benchmarking web service that enables researchers to compare scene graph generation methods and increase the visibility of new methods in a central place.
The paper also includes a comparison of existing panoptic scene graph generation methods using the discussed metrics, providing insights into the performance of these models.
To Another Language
from source content
arxiv.org
Djupare frågor