Dynamical Recurrent Visual Perceiver

Real-life applications of computer vision often require high-quality visual sensors. With current technology, such sensors are expensive.

Empirical evidence from our ERC-funded research of biological vision suggests that eye motion enhances recognition capabilities beyond what could be expected if the eye was functioning as a static camera.

Motivated by these findings, here we aim to perform the major computer vision tasks required for many real-life applications, including segmentation, classification and identification, with low-resolution cameras.

The idea is to use a series of low-resolution frames from a moving camera, rather than using a single high-resolution image. This novel algorithm, termed here DRVis, can be implemented in software or hardware.

Unlike existing solutions that use multiple frames to reconstruct a high resolution image from low resolution ones, DRVis does not need to learn all the particularities needed for reconstruction; instead, it focuses on extracting the necessary features per the given task.

The goals of the PoC project are to scale up and diversify our current software system, implement the system in hardware and demonstrate its performance in field conditions and develop the IPR strategy and explore the commercialization potential of our solution.

We expect it will be applicable to a wide spectrum of image processing tasks in settings where sensor quality is low but multiple time samples are available including smart agriculture, drone navigation and visual aid devices.