iFRAME (inhomogeneous Filters Random Field And Maximum Entropy)

Experiment 3.3: Clustering by fitting Mixture of iFRAME models without local shift (DoG version)

Clustering by EM-like algorithm by fitting Mixture of iFRAME Models. The clustering is model-based, which does NOT require or rely on pairwise similarity meausre. The clustering uses hard classification.
Three characteristics: (1) Multiple scales of Gabor features and DoG for good representation; (2) Multiple chains of HMC sampling for accurate parameters estimation, (3) Multiple GPUs of parallel computating for fast computation.

Experiment 3.1: hard EM version
Experiment 3.2: soft EM version
Contents
Case 1: Horses facing two different directions
Case 2: Horses and zebras
Case 3: Butterflies, dragonflies, and bugs
Case 4: Human faces, butterflies, and horses
Case 5: Cars, motorbikes, scooters, and bikes

Case 1: Horses facing two different directions (details)

templates





Synthesized templates by multiple chains

Case 2: Horses and zebras (details)

templates




Synthesized templates by multiple chains

Case 3: Butterflies, dragonflies, and bugs (details)

templates





Synthesized templates by multiple chains

Case 4: Human faces, butterflies, and horses (details)

templates




Synthesized templates by multiple chains

Case 5: Cars, motorbikes, scooters, and bikes (details)

templates






Synthesized templates by multiple chains