Hyperfeatures – Multilevel Local Coding for Visual Recognition

Abstract

Abstract. Histograms of local appearance descriptors are a popular representation for visual recognition. They are highly discriminant and have good resistance to local occlusions and to geometric and photometric variations, but they are not able to exploit spatial co-occurrence statistics at scales larger than their local input patches. We present a new multilevel visual representation, 'hyperfeatures', that is designed to remedy this. The starting point is the familiar notion that to detect object parts, in practice it often suffices to detect co-occurrences of more local object fragments -a process that can be formalized as comparison (e.g. vector quantization) of image patches against a codebook of known fragments, followed by local aggregation of the resulting codebook membership vectors to detect cooccurrences. This process converts local collections of image descriptor vectors into somewhat less local histogram vectors -higher-level but spatially coarser descriptors. We observe that as the output is again a local descriptor vector, the process can be iterated, and that doing so captures and codes ever larger assemblies of object parts and increasingly abstract or 'semantic' image properties. We formulate the hyperfeatures model and study its performance under several different image coding methods including clustering based Vector Quantization, Gaussian Mixtures, and combinations of these with Latent Dirichlet Allocation. We find that the resulting high-level features provide improved performance in several object image and texture image classification tasks.

Information Theoretic Key Frame Selection for Action Recognition

“…However, these approaches lack the relations between the features in the spatial and the temporal domains which are helpful for recognition. There are many recent research on extending "bag-of-words" to add the spatial relation in the context of object categorization [17,1,11,8,10]. In particular, pyramid match kernel [8,10] used the weighted multi-resolution histogram intersection as a kernel function for classification with sets of image features.…”

Discovery of Collocation Patterns: from Visual Words to Visual Phrases

“…Many existing work devote to seeking visual features that are spatially co-occurrent, such as "semi-local" parts [12], depedency regions [22], frequent spatial configurations [16], perceptual groups of local features [10], sparse flexible model [2] and hyperfeatures [1]. Various criteria are proposed to measure the spatial dependency among the primitive visual features.…”

Spatiotemporal pyramid representation for recognition of facial expressions and hand gestures

“…These are important correlation, which are helpful for recognition. There are many recent research on extending "bag-of-words" to add the spatial relation in the context of object categorization [15,1,10,7,9]. In particular, pyramid match kernel [7] used the weighted multi-resolution histogram intersection as a kernel function for classification with sets of image features.…”

“….00 .00 .00 .00 1.0 .00 .00 .00 .00 .00 .00 1 93.50 pLSA [20] 50.00 pLSA-ISM [20] 83.33 Table 3. The facial expression recognition rates obtained from different algorithms with leave-one-out cross-validation experiment setting.…”