Buffl

Q8

D
by David

Torralba and Oliva published a paper entitled "Statistics of natural image categories". They report that, contrary to what was believed before, the    ---- bitte auswählen ----  power spectrum  pixel intensity histogram   of (images of ) natural scenes is only    ---- bitte auswählen ----  non-isotropic  isotropic   if averaged across image categories, but if analysed separately for different image categories, they found strong correlations between   ---- bitte auswählen ----  the shape  the total power  phase  the complex conjugate   of the power spectrum and image categories. Typically a density-plot the power spectrum of an image of a man-made scene is more    ---- bitte auswählen ----  egg-shaped  circular   triangluar-shaped  star-shaped  . Based on   ---- bitte auswählen ----  two  a large number of   a small number of  a single   component(s) of a principal component analysis (PCA) performed on the power spectrum, Torralba & Oliva were able to correctly categorise images into animal and non-animal scenes in % of the cases. Calculating the PCA of the power spectrum is a   ---- bitte auswählen ----  non-linear  linear   operation.    ---- bitte auswählen ----  However  Still    this operation could be performed   ---- bitte auswählen ----  only with great difficulty  already in the retina   in a feedforward manner  only using feedback   in the human brain given what is currently known about physiology. Thus Torralba & Oliva concluded that animal versus non-animal categorization is so rapid because their   ---- bitte auswählen ----  structural description model  summary statistic  image segmentation  view-based   approach does not require an explicit    ---- bitte auswählen ----  image segmentation  Fourier transformation  image alignement   step.

Torralba and Oliva published a paper entitled "Statistics of natural image categories". They report that, contrary to what was believed before, the  power spectrum of (images of ) natural scenes is only  isotropic if averaged across image categories, but if analysed separately for different image categories, they found strong correlations between the shape of the power spectrum and image categories. Typically a density-plot the power spectrum of an image of a man-made scene is more  star-shaped . Based on a small number of component(s) of a principal component analysis (PCA) performed on the power spectrum, Torralba & Oliva were able to correctly categorise images into animal and non-animal scenes in 80 % of the cases. Calculating the PCA of the power spectrum is a non-linear operationStill this operation could be performed in a feedforward manner in the human brain given what is currently known about physiology. Thus Torralba & Oliva concluded that animal versus non-animal categorization is so rapid because their  summary statistic approach does not require an explicit  image segmentation step.

Thorpe, Fize & Merlot published a study in Nature which exerted a very strong influence on the object recognition community. In their paper they showed that   ---- bitte auswählen ----  human observers  monkeys   cats   could decide whether a previously unseen   ---- bitte auswählen ----  photograph  line drawing  painting   of a natural scene contained an animal or not. The median reaction time (RT) of the observers was around   ---- bitte auswählen ----  400-500  100-200  200-300  500-600  600-700   ms with a mean percentage correct of    ---- bitte auswählen ----  85-90  90-95  95-100  80-85   % correct (note that the observers showed   ---- bitte auswählen ----  a slight  no  a strong   speed-accuracy trade-off). Subsequent   ---- bitte auswählen ----  ERP  fMRI  PET  multi-unit  single-unit   analyses showed that roughly   ---- bitte auswählen ----  150  100  200  250   ms after stimulus onset the measured neurophysiological correlate could already reliable signal the presence or absence of an animal in a post-hoc analysis. Thus processing of the natural scene stimulus was already completed after such a comparatively short time. According to the authors this result provides strong evidence in favour of essentially    ---- bitte auswählen ----  feedforward  dynamic feedback  multi-level feedforward & feedback  feedback  deep-belief neural network    theories of visual object recognition. This, in turn, argues against object recognition theories requiring an explicit    ---- bitte auswählen ----  image segmentation  2D-to-3D  Fourier transform  Wavelet transform  multi-scale image decomposition   step prior to recognition, as such a step is presumed to require   ---- bitte auswählen ----  time consuming  fast  computationally complex     ---- bitte auswählen ----  iterative  feedback  feedforward  non-linear processing  linear decomposition  fast Fourier   algorithms.

Thorpe, Fize & Merlot published a study in Nature which exerted a very strong influence on the object recognition community. In their paper they showed that human observers could decide whether a previously unseen photograph of a natural scene contained an animal or not. The median reaction time (RT) of the observers was around 400-500 ms with a mean percentage correct of  90-95 % correct (note that the observers showed  a slight speed-accuracy trade-off). Subsequent ERP analyses showed that roughly 150 ms after stimulus onset the measured neurophysiological correlate could already reliable signal the presence or absence of an animal in a post-hoc analysis. Thus processing of the natural scene stimulus was already completed after such a comparatively short time. According to the authors this result provides strong evidence in favour of essentially  feedforward theories of visual object recognition. This, in turn, argues against object recognition theories requiring an explicit  image segmentation step prior to recognition, as such a step is presumed to require  time consuming iterative/feedback (both correct as it seems) algorithms.

Wichmann, Drewes, Rosas and Gegenfurtner published a paper   ---- bitte auswählen ----  casting doubt on  confirming   the conclusions made by Torralba & Oliva. The two main conclusions of the study – based on   ---- bitte auswählen ----  computational analysis  psychophysical experiments  neuro-imaging techniques   – were, first, that for human observer animal detection in typical photographs of natural scenes   ---- bitte auswählen ----  is independent of the power spectrum  relies on many PCA components of the power spectrum  depends on the power spectrum as claimed by Torralba & Oliva  is independent of the phase spectrum  . Second, they may indicate that in typical, commercial databases the statistics of the images may   ---- bitte auswählen ----  be as  not be as  even be more   natural as/than often presumed, because photographs typically represent a   ---- bitte auswählen ----  true random sample  Gaussian sample  biased  unbiased   view of the world.

Wichmann, Drewes, Rosas and Gegenfurtner published a paper  casting doubt on the conclusions made by Torralba & Oliva. The two main conclusions of the study – based on psychophysical experiments – were, first, that for human observer animal detection in typical photographs of natural scenes is independent of the power spectrum . Second, they may indicate that in typical, commercial databases the statistics of the images may not be as natural as/than often presumed, because photographs typically represent a biased view of the world.

Author

David

Information

Last changed