SynthRel0: Towards a Diagnostic Dataset for Relational Representation Learning

Daniel Dorda, Moin Nabi; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 0-0

This work analyses the sources of complexity in scene graph proposal problems, and develops a mathematical framework for efficiently designing synthetic relationship models. An entropy based metric is proposed for measuring the ambiguity of relational datasets. Using these tools, a first approximation to a synthetic dataset is given, and experiments with a simple baseline are performed to show how the difficulty of the proposed task changes with varying dataset parameters, like missing annotation ratio and feature granularity. These experiments illuminate the desirable qualities of future synthetic relationship datasets.