Integration of these new dynamic sensory modules in a visual architecture and experimental study of their performances in the case of degraded visual sources.
Objective: Validate the non-standard bio-inspired early-vision front-end on realistic data set, targeting low vision (e.g. underwater) applications.
Methods: Since a new innovative early-vision front-end is going to be made available thanks to the previous tasks, the final step is to validate it embedded in a larger biologically inspired visual system and targeting an application for which state of the art visual methods partially fail.
In addition to this benchmarks, the non-standard bio-inspired early-vision front-end results are going to be compared to high-level machine learning mechanisms, e.g. as novelty detectors (Kassab et al, 2009), in order to quantify the obtained performances.
Since we want to experiment to which extent such an improved early-vision front-end contribute to visual perception, we are going to not only take basic visual cues detection into account, but to experiment on high-level visual functions. Concrete demonstration of cognitive tasks enhancement are going to include:
(i) Non-linear static/dynamic cues detection: calculation of maps of e.g. colored- texture / background-motion / object-motion with segmentation of uniform regions wr.t. the cue.
(ii) Gesture recognition: discriminate between two different displacements (e.g. walk versus march, or crowd behaviour).
(iii) Image category recognition: recognize the image general category (e.g. a natural versus artificial scene, an animal versus a manufactured object).
(iv) Detection of unexpected event: recognize an unexpected displacement (i.e. a failure of prediction in local motion detector).
(v) Image segmentation from categorization: when a categorization is performed the retinal units with a non-negligible contribution to this categorization process provide a cue about the part of the image corresponding to this categorization, thus allow the segmentation.
This means that, at the biologically inspired modelling model level, extra-cortical functions in connection with non-standard retinal cells response are going to be studied, including focus of attention and motivated vision.
Task steps:
(i) Specifications of a set of benchmark validation tests (object category recognition, novelty detection).
(ii) Deployment of the benchmarking platform data and software.
(iii) Realization of a benchmarking test set against general non-biologically constrained algorithms.
