Citation Context

...s that the PHOW features, efficiently computed by using VLFeat, are a state-of-the-art image descriptors (better classification results can be obtained by combining a variety of different descriptors =-=[5]-=-). 5. DESIGN Granularity. VLFeat avoids implementing specific applications (e.g. there is no face detector); instead it focuses in cimplementing algorithms (feature detectors, clustering, etc.) that c...

Citation Context

...riptors on the same dataset are shown on Table 2. Note that better performance has been reported with multiple descriptor types (e.g., methods using multiple kernel learning have achieved 77.7% ± 0.3 =-=[7]-=- and 78.0% ± 0.3 [28, 2] on Caltech-101 with 30 training examples), or subcategory learning (83% on Caltech101 [26]). The coding and pooling module combinations used in [27, 31] are included in our co...

Citation Context

...am of visual words is stored, d is the minimum of the number of words detected per image and the vocabulary size. For a GIST descriptor [19], d is of the order of 1000. For multiple-kernel techniques =-=[6]-=-, d might be of the order of 20, 000. For the system in this paper, d can be as low as 1500, while still leveraging all the descriptors used in the multiple-kernel technique. Note that although we sha...

Citation Context

...BoW or histogram of oriented gradient (HOG) [18, 4] features. In the current state-of-the-art, multiple descriptors and kernels are combined using either ad hoc or multiple kernel learning approaches =-=[19, 5, 20, 21]-=-.What does classifying more than 10,000 image categories tell us? 3 Work in machine learning supports using winner-takes-all between 1-vs-all classifiers for the final multi-class classification deci...

Citation Context

...rs in a square image. 3 Please refer to supplementary material for illustrative videos. Such an approach can be considered a basic form of multiple kernel learning (MKL), and indeed it has been shown =-=[9]-=- that in terms of performance full MKL (in which the relative weighting of the different kernels is learned from training data) does not provide a great deal of improvement over this simple approach. ...

Citation Context

...ond method, which employs a joint formulation for learning the global weights for feature channels. This global weighting is analogous to learning the kernel weights in multiple kernel learning (MKL) =-=[11]-=-. In MKL, the task is to select informative kernels, whereas here we try to select informative feature channels. We formulate the optimization as follows: simply assign the correspondingm(B f i min w,...

Citation Context

...rns. Match kernels defined over various pixel attributes provide a unified way to generate a rich, diverse visual feature set, which has been shown to be very successful to boost recognition accuracy =-=[6]-=-. As validated by our own experiments, gradient, color and shape match kernels are strong in their own right and complement one another. Their combination turn out to be always (much) better than the ...

Citation Context

...ors for object categorization. Our experimental results on some benchmark data sets show that MTJSRC is competitive to several state-of-the-art multiple kernel learning methods for object recognition =-=[25]-=-–[27]. 2) Face recognition for dynamic videos. In this application, we assume that each video sequence contains only one single subject. Given an input video, we first employ a face detector [28] to l...

Citation Context

...formance has been reported with color images (e.g., 78.5% ± 0.4 with a saliency-based approach [22]), multiple descriptor types (e.g., methods using multiple kernel learning have achieved 77.7% ± 0.3 =-=[12]-=-, 78.0% ± 0.3 [2, 37] and 84.3% [40] on Caltech-101 with 30 training examples), or subcategory learning (83% on Caltech-101 [35]). On the Scenes benchmark, preclustering does improve results for small...

Citation Context

...efficient learning and test evaluation. The current best published results on Caltech256 were obtained by a kernel combination classifier using 39 different feature kernels, one for each feature type =-=[13]-=-. However, since both training as well testing are computationally very expensive with this classifier, this model is unsuitable for our needs. 4Instead, in this work we use as image representation t...