80 likes | 176 Views
Adam Yeh. UCF Computer Vision REU Week 4. Coding Problems. Problem: Adaboost taking too much memory: 160k features, feature value as int (4 bytes) Maximum heap size: 1.5GB Each picture takes 160k*4=640kB ~2000 pictures can be trained on Solution: store less information
E N D
Adam Yeh UCF Computer Vision REU Week 4
Coding Problems • Problem: Adaboost taking too much memory: • 160k features, feature value as int (4 bytes) • Maximum heap size: 1.5GB • Each picture takes 160k*4=640kB • ~2000 pictures can be trained on • Solution: store less information • Calculate values as needed, less preprocessing
Feature Selection • Previously: Haar basis wavelets • Pixel sums and differences [1]
Feature Selection • Gaussian Derivatives: Silapachote et al. • Calculate with varying sigma, location, and extract features
Feature Selection • Gabor Filters • Harmonic (sin/cos) function multiplied by Gaussian • Use different frequencies, orientations
Progress: This Week • Adaboost code w/Haar features • Still training on 1500/4500 • Started coding for Gaussian features • Obtained test images (MMI)
Next Week • Test images: MMI Face DB • 800+ videos, 200+images • 8GB+ • Need to download, process, sort/organize • Other features • Coding/testing Gaussian, Gabor features • Attentional Cascade?
References • [1] Viola-Jones • [2] Silpachote, P, Karuppiah, D, and Hanson, A. “Feature Selection Using Adaboost for Face Expression Recognition”. Proceedings of the Fourth IASTED International Conference: Visualization, Imaging, and Image Processing, Sept 2004. • [3] Wikipedia