The GENOA
|
![]() |
- Abstract - GENoa hUman Active
fixation database:PEripersonal space STereoscopic images and grOund
truth disparity GENUA PESTO is a large database of stereoscopic images (2700) obtained mimicking the human eye posture, i.e. including vergence and cyclotorsion. The database provides the stereoscopic pairs together with the ground-truth information of:
|
![]() |
![]() |
![]() |
Cortical
|
![]() |
- Abstract - The organization of visual pathway and
especially of striate cortex, suggests that the structural principles of
neural circuitry (e.g., inhibitory mechanisms, feedback loops, clustered
connectivity, cortical modularity, etc.) might be a powerful computational
solution. By studying the basic mechanisms and circuits available for biological
visual processing, one can obtain powerful computational solutions for
complex visual tasks in machine vision. Specifically, we focus our attention
on:
|
![]() |
![]() |
![]() |
Motion
|
![]() |
- Abstract - Motion in depth can be estimated by evaluating
the time derivative of local stereo disparity. Starting from phase-based
approaches to disparity estimation, we propose a computational method to
obtain a stable and dense map of velocity-in-depth. On the same theoretical
considerations, we are working on the definition of a space-time neuromorphic
filtering model of binocular cells selectively responsive to motion in
depth.
|
![]() |
![]() |
![]() |
Analysis of
|
![]() |
- Abstract - Under the hypothesis of an isomorphic
(i.e., ``physicalist'') MT representation of local velocity, based on traveling
waves of activation, we are investigating the functionality of a simple
model for MST cells. The sensitivity of such MST cells to specific optical
flow components/patterns (e.g., expansion, contraction, rotation, looming
) is obtained through spatio-temporal integration over MT cortical activity
instead by vector operations. Implications to the analysis of optic flow
fields originated from egomotion are investigated.
|
![]() |
![]() |
![]() |
Discrete
|
![]() |
- Abstract - In real world situations, many alerting
and guidance tasks can be performed by peripheral vision on the base
of dynamic measurements over gross-feature space, without a detailed image
understanding. In this context, the development of analog VLSI sensory
modules with embedded dynamic perceptual processing is particularly appealing
to build behavioural systems able to provide a reflex as a shortcut to action.
The computational principles of these dedicated analog VLSI modules are studied
on
the assumption that a most efficient implementation of (preattentive) perceptual tasks can be based on analog structured arrays of simple interacting units (lattice networks) reacting collectively to spatio-temporal input stimuli. The topological organization of the interaction among units determines their functional response. It has been observed that each type of lattice network can be sustained by a unified entity, or computational engine, on which the various perceptual algorithms can be mapped by acting properly onthe pattern of afferent connections. An interesting case study is represented by an 1D diffusive layer with linear next-nearest recurrent inhibitory reactions which leads to periodic kernels similar to Gabor functions. The circuital interpretation of these lattice networks as a mesh of one-way interacting elements (e.g., current controlled current sources) led to efficient aVLSI solutions toward compact optical microsystems which combine sensorial and computational capabilities.
|
![]() |
![]() |
![]() |
Coupled
|
![]() |
- Abstract - The computational capabilities of 2D
continuous distribution of locally-interacting neural cells are analyzed
through nonlinear partial derivative equations of reaction-diffusion type.
The resulting coupled diffusion maps can be interpreted as a phenomenological
model for the processes that take place on visual cortical surface. Averaging
the pattern of activity over a window comparable in size with those of
a complete set of orientations, we gain instantaneous global perception
supporting scene segmentation. Experiments on texture segmentation, conducted
with artificial and natural textures, validates the approach.
|