Gaussian Mixture Modeling (GMM) és Fisher vector toolkit
News:
CLIFF v0.11:
=========
Open Compute Library Image Feature Framework (CLIFF) is an OpenCL based image feature framework to compute point descriptors and modeling images for classification and retrieval efficiently on GPU-s and multicore CPU-s
CLIFF features
- local feature extraction
- local, dense and spatial Fisher vector [1]
- soft cluster assigment
- spatial [4] and segmentation based pooling
- PCA and GMM models for HOG[2] and Color moments
Results on Pascal VOC 2007 dataset:
MAP 0.6415, ColHOG, GMM with 512, spatial pooling (1x1, 3x1)
Used options:
- -pca models/pca.clch
- -gmm models/gmm.clch.512
- -step 8
- -scales 4
- -spatial_pool
- -upscale
- -pyr 5
- -colhog
- -alpha 0.125
- -norm 2
- System requirements :
- - x86 compatible Intel/AMD processor
- - Linux with AMD APP SDK or Mac OS X 10.7+
- - gcc-4.3
- - OpenCL 1.2
- - OpenCV 2.1+
- - pkg-config
- dimension reduction via PCA
- local Fisher vector for low-level features
- spatial [4] and segmentation based pooling
- PCA and GMM models for LBP[5], SIFT[3], HOG[2] and Color moments
- System requirements :
- - x86 compatible Intel/AMD processor
- - gcc-4.3
- - Cuda 1.0 compute compatibly GPU (see CUDA programming guide : http://developer.nvidia.com/object/gpucomputing.html)
- - Nvidia CUDA SDK 1.3
References
- F. Perronnin and C. Dance. Fisher kernels on visual vocabularies for image categorization. In IEEE Conference on Computer Vision and Pattern Recognition, 2007. CVPR’07, pages 1–8, 2007.
- Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. CVPR 2005
- D.G. Lowe. Object recognition from local scale-invariant features. In International Conference on Computer Vision, volume 2, pages 1150–1157, 1999.
- C. Schmid S. Lazebnik and J. Ponce. Beyond Bags of Features: Spatial Pyramid Matching for Recognizing Natural Scene Categories. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, New York, June 2006, 2006.
- T. Ojala, M. Pietikäinen, and D. Harwood (1996), "A Comparative Study of Texture Measures with Classification Based on Feature Distributions", Pattern Recognition, vol. 29, pp. 51-59.
Contact
Bálint Daróczy
MTA SZTAKI, Data Mining and Search Group
Lágymányosi str. 11, Budapest, Hungary
daroczyb_at_ilab.sztaki.hu
CLIFF alpha Rev. 0.11 (see README in the compressed file)
CUDA based Gaussian Mixture Modeling (GMM) and Fisher vector toolkit:
=============================================================
Rev. 0.94 (see README in the compressed file) Download:
This implementation of Gaussian Mixture Model using EM algorithm is an efficient GPU based tool to train GMM models on large dimensional data (100+) within a short time. For details see our report. The source code was implemented fully in C++ compiled with nvcc and gcc-4.3.2. The compressed file also includes a tool and model files to create Fisher vectors[1] from image patches such as HOG[2] or SIFT[3]
Fisher vector Toolkit features: