Point-based matching applied to images generated by log(z) and log(z+alpha) forms of artificial retina

We present the results of an investigation that compares matching SIFT local image features that have been extracted from a software-based retina model under two different forms of retina geometry, log(z) and log(z + α). Our retinas are sampled by receptive fields (RF) which are organized at a high density in the central foveal region of the retina and with progressively sparser density in the surrounding periphery in a configuration similar to that found in biological vision. We demonstrate our retinas using either point-based or variable-sized overlapping Gaussian kernel sampling of the input image and present visualisation results depicting the sampled retina responses back-projected as images in retinal coordinates. Multi-resolution, space-variant visual information is extracted on a scale-space continuum and SIFT interest point descriptors are extracted from these back-projected images in order to represent the visual appearance of local regions surrounding SIFT interest points. This paper describes the design, implementation and initial evaluation of space variant artificial log(z) and log(z+α) retina tessellations comprising a circular overlapping RF model. We compare the matching performance of the back-projected space-variant response using standard SIFT by plotting receiver operating characteristic (ROC) curves. While the primary objective of retina sampled SIFT is to reduce feature data rates while focusing attention in the context of visual search, our preliminary matching results indicate that SIFT applied to the log(z + α) retina images outperformed SIFT applied to the log(z) retina by 5% and 4% at false alarm rates set to 10% and 20% respectively.