Beyond Photo-Domain Object Recognition: Benchmarks for the Cross-Depiction Problem

Abstract

The cross-depiction problem is that of recognising visual objects regardless of whether they are photographed, painted, drawn, etc. It introduces great challenge as the variance across photo and art domains is much larger than either alone. We extensively evaluate classification, domain adaptation and detection benchmarks for leading techniques, demonstrating that none perform consistently well given the cross-depiction problem. Finally we refine the DPM model, based on query expansion, enabling it to bridge the gap across depiction boundaries to some extent.