Classification Regardless of Depictive Style


Humans are able to recognise objects across a remarkably wide variety of depictive styles. Whether a horse is seen for real, photographs, drawn, painted, sculpted in bronze or wire or glass, or as a wooden toy. In contrast, computers exhibit a far more limited capacity to recognise regardless of depiciton.

This project is funded by EPSRC, grant Classification Regardless of Depictive Style, EP/K015966/1.

Prime Shapes: We have discovered that up to 80% of segments of real world images naturally classify into one a few shapes.  Applications include scene identification and qualitative shape as a feature.Prime_Shapes.html
Visual Class Models: Visual class models are constructed by using  prime shapes to label nodes in graph; both shapes and graph are automatically learned from input examples. Performance rates exceed standard Bag of Words with regards to cross-depiction classification.Visual_Class_Models.html