Картинки на тему «Improving the Fisher Kernel for Large-Scale Image Classi?cation» |
| Курсы английского | ||
| << Create the Internet of Your Things: The Microsoft Vision for IoT | NARUC Staff Subcommittee Accounting and Finance Accounting for Emission Allowances >> |
Автор: Lempitsky. Чтобы познакомиться с картинкой полного размера, нажмите на её эскиз. Чтобы можно было использовать все картинки для урока английского языка, скачайте бесплатно презентацию «Improving the Fisher Kernel for Large-Scale Image Classi?cation.ppt» со всеми картинками в zip-архиве размером 852 КБ.
Improving the Fisher Kernel for Large-Scale Image Classi?cation
содержание презентации «Improving the Fisher Kernel for Large-Scale Image Classi?cation.ppt»| Сл | Текст | Сл | Текст |
| 1 | Improving the Fisher Kernel for | 8 | and Dance//CVPR07] suggest a diagonal |
| Large-Scale Image Classi?cation. Florent | approximation to Fisher matrix: | ||
| Perronnin, Jorge Sanchez, and Thomas | 9 | Classification with Fisher kernels. | |
| Mensink, ECCV 2010. VGG reading group, | Use whitened Fisher vectors as an input to | ||
| January 2011, presented by V. Lempitsky. | e.g. linear SVM Small codebooks (e.g. 100 | ||
| 2 | From generative modeling to features. | words) are sufficient Encoding runs faster | |
| dataset. Discriminative classfier model. | than BoW with large codebooks (although | ||
| Input sample. Generative model. fitting. | with approximate NN this is not so | ||
| Parameters of the fit. | straightforward!) Slightly better accuracy | ||
| 3 | Simplest example. Dataset of vectors. | than “plain, linear BoW”. F. Peronnin and | |
| Discriminative classfier model. Input | C. Dance // CVPR 2007. | ||
| vector. K-means. Codebook. fitting. | 10 | Improvements to Fisher Kernels. =0. | |
| Codebooks Sparse or dense component | Perronnin, Jorge Sanchez, and Thomas | ||
| analysis Deep belief networks Color GMMs | Mensink, ECCV 2010. Overall very similar | ||
| .... Closest codeword. | to how people improve regular BoW | ||
| 4 | Fisher vector idea. Jaakkola, T., | classification. Idea 1: normalization of | |
| Haussler, D.: Exploiting generative models | Fisher vectors. Justification: our GMM. | ||
| in discriminative classi?ers. NIPS’99. | probability distribution of VW in an | ||
| Generative model. Discriminative classfier | image. Assume: Image specific “content”. | ||
| model. Input sample. Parameters of the | Then: Thus: Observation: image | ||
| fit. fitting. Information loss (generative | non-specific “content” affects the length | ||
| models are always inaccurate!). Can we | of the vector, but not direction. | ||
| retain some of the lost information | Conclusion: normalize to remove the effect | ||
| without building better generative model? | of non-specific “content” ...also | ||
| Main idea: retain information about the | L2-normalization ensures K(x,x) = 1 and | ||
| fitting error for the best fit. Same best | improves BoV [Vedaldi et al. ICCV’09]. | ||
| fit, but different fitting errors! | 11 | Improvement: power normalization. ? | |
| 5 | Fisher vector idea. X. ? (?1,?2). | =0.5 i.e. square root works well c.f. for | |
| Fisher vector: Jaakkola, T., Haussler, D.: | example [Vedaldi and Zisserman// CVPR10] | ||
| Exploiting generative models in | or [Peronnin et al.//CVPR10] on the use of | ||
| discriminative classi?ers. NIPS’99. | square root and Hellinger’s kernel for | ||
| Generative model. Discriminative classfier | BoW. | ||
| model. Input sample. Fisher vector. | 12 | Improvement 3: spatial pyramids. Fully | |
| fitting. Main idea: retain information | standard spatial pyramids [Lazebnik et | ||
| about the fitting error of the best fit. | al.] with sum-pooling. | ||
| 6 | Fisher vector for image | 13 | Results: Pascal 2007. Details: regular |
| classification. F. Peronnin and C. Dance | grid, multiple scales, SIFT and local RGB | ||
| // CVPR 2007. Assuming independence | color layout, both reduced to 64 | ||
| between the observed T features Encoding | dimensions via PCA. | ||
| each visual feature (e.g. SIFT) extracted | 14 | Results: Caltech 256. | |
| from image to a Fisher vector Using | 15 | PASCAL + additional training data. | |
| N-component gaussian mixture models with | Flickr groups up to 25000 per class | ||
| diagonalized covariance matrices: N | ImageNet up to 25000 per class. | ||
| dimensions. 128N dimensions. 128N | 16 | Conclusion. Fisher kernels – good way | |
| dimensions. | to exploit your generative model Fisher | ||
| 7 | Relation to BoW. BoW. Extra info. F. | kernels based on GMMs in SIFT space lead | |
| Peronnin and C. Dance // CVPR 2007. N | to state-of-the-art results (on par with | ||
| dimensions. 128N dimensions. 128N | the most recent BoW with soft assignments) | ||
| dimensions. | Main advantage of FK over BoW are smaller | ||
| 8 | Whitening the data. Fisher matrix | dictionaries ...although FV are less | |
| (covariance matrix for Fisher vectors): | sparse than BoV Peronnin et al. trained | ||
| Whitening the data (setting the covariance | their system within a day for 20 classes | ||
| to identity): Fisher matrix is hard to | for 350K images on 1 CPU. | ||
| estimate. Approximations needed: [Peronnin | |||
| Improving the Fisher Kernel for Large-Scale Image Classi?cation.ppt | |||
Improving the Fisher Kernel for Large-Scale Image Classi?cation
«The animals» - FISH. KOALA. FLAMINGO. EMU. The animals which live in the polar regions. The animals which live in the forest. ZEBRA. The animals which live in the OCEAN. SEA-HORSE. STARFISH. PENGUIN. The animals which live in a SAVANNA. PANDA. KANGAROO. The animals which live in the rainforest and tropics. SNAKE. TIGER.
«English for you» - EuroTalk. Что ты научишься делать и узнаешь. Ты убедишься насколько интересным и захватывающим может быть обучение языку. Ты научишься правильно строить предложение. Узнать насколько хорошо ты усвоил материал тебе помогут: Слова и выражения по темам. ENGLISH FOR YOU. Все слова и выражения озвучены носителями языка.
«The english-speaking countries» - Australia. The English-speaking countries. Scotland. USA. Great Britain. Disneyland.
«Женщина the woman» - Наименование молодой девушки в современном английском языке. «Un homme»- франц. « A man »- англ. Женский интеллект. The wife is the key to the house. Chicken’s mind- Куриные мозги. As great a pity to see a woman cry as a goose go barefoot. Оценочная структура лексической единицы “женщина”. Баба слезами беде помогает.
«The green movement» - The main objective — to achieve the decision of global environmental problems, including by attraction to them of attention of the public and the authorities. Management of each such branch in which head there is a chief executive, carries out national board. It became the first African who has headed this organization.
«Английский Cambridge» - www.cambridge-centre.ru. Сейчас планирую подготовиться и получить еще сертификат по финансовому английскому. Короткие тренинги для специалистов и компаний. Тема «Телефонные переговоры». Вопрос 6. Где проходят занятия? Минимальная продолжительность курса-тренинга 8 академических часов. О глобализации, интеграции, слиянии, поглощении…
