Learning Ordinal Relationships for Mid-Level Vision

D. Zoran, P. Isola, D. Krishnan and W. T. Freeman, to appear in ICCV 2015, Santiago, Chile

teaser

Abstract: We propose a framework which infers mid-level visual properties of an image by learning about ordinal relationships. Instead of estimating metric quantities directly, thesystem proposes ordinal relationship estimates for pairs of points in the input image. These probabilistic ordinal measurements are then aggregated and globalized to create a full output map of continuous metric measurements. Estimating order relationships between pairs of points has several advantages over metric estimation: it requires solving a simpler three way classification problem instead of metric regression; humans are better at making relative judgements so data collection is easier, and ordinal relationships are invariant to monotonic transformations of the data, thereby increasing the robustness of the system and providing qualitatively different information. We demonstrate that our framework works well on two important mid-level vision tasks: intrinsic image decomposition and depth from a single RGB image. We train two separate systems with the same architecture on data from two different modalities. We provide an analysis of the resulting models, showing that they learn a number of simple rules to make ordinal decisions. Finally, we show that we are able to achieve competitive results on depth estimation and state-of-the-art results on intrinsic image decomposition.

Paper

Code

IIW Test Image Data