1Book Cover RecognitionSHAO-CHUAN (SHAWN) WANGCMOS E6737 Biometrics Final Projectsw2644@columbia.eduhttp://shaochuan.info/
Book Cover RecognitionSystem Goal:Input: A middle resolution image taken from web camera or cell phone (640 x 480) that contains a book.Output: The book title/id.   2Book CoverRecognition System“Learning from Data”
Baseline Model: Bag-of-wordsLow level feature extraction (Dense SIFT) (CVPR09)Visual words learning (a.k.a codebook learning) via vector quantization.Spatial pooling of local features. Linear SVM classification. (libLinear package)3
A Glance on Dataset (1/2)9 books. training on half (15), test on half (15):4
A Glance on Dataset (2/2)5
Baseline: ExperimentsSystem parameters:15 training images per book, test on the rest.Resize image size to 300 pixels (long side)Visual codebook size (K = 225)# of spatial pyramid level = 0,1,2Spatial pooling function = maxL2-regularized 1-norm loss linear SVM5-fold cross validationResults:Nearly perfect recognition results. (99.259%=134/135)6
More Challenging Testing ImagesMore clutter, more occlusion, more realistic.7
More Challenging: ExperimentsResults:Poor: 36.24% (MAP) 8Confusion matrixSo, Can we “detect” and “crop” books automatically?
Hough TransformAssumptions:Books are rectangular, some contrasts between books and clutter.No many “false” strong edges on clutter.9
Auto cropping results (1/2)10Easier cases:
Auto cropping results (2/2)Really difficult cases:11
More Challenging: ExperimentsResults:Without cropping: 36.24% (MAP)With autocropping: 58.60% (MAP)12
Book Cover Recognition

Book Cover Recognition

Editor's Notes

  • #2 Left to the title, a presenter can insert his/her own image pertinent to the presentation.
  • #14 The End of Slideshow.