STSIM-C Texture Similarity

Experimental Procedure

0. STSIM vector extraction
1. Compute covariance matrix
2. Approximate Cross Validation
   • Compute cluster center or set example for retrieval
   • Measure Precision@1 and RoC

Experimental Settings:

• Global or Intraclass covariance matrix calculation
• Color or Grayscale features

Datasets:

• CuReT: Fewer classes, many examples per class
  • Represent class with many exemplars
  • Tests under variation of lighting and viewing conditions
• Identical: Same texture, perceptually identical
• ViSiProg: Different classes; Perceptually similar
• Distortions: Single image with applied distortion translation, rotations, scaling, etc.

Differences from previous work:

• Previous work would have made assumptions among feature correlation
• Similarity is order invariant
• 1-,2-Color improves accuracym; Minimal gains for 4-color