Light-weight semantic segmentation for compound images

The eye structure of insects, which is called a compound eye, has interesting advantages. It has a large field of view, low aberrations, compact size, short image processing time, and an infinite depth of field. If we can design a compound eye camera which mimics the compound eye structure of insects, compound images with these interesting advantages can be obtained. In this paper, we consider the design of a compound camera prototype and low complexity semantic segmentation scheme for compound images. The prototype has a hemisphere shape and consists of several synchronized single-lens reflex camera modules. Images captured from camera modules are mapped to compound images using multi-view geometry to emulate a compound eye. In this way, we can simulate various configurations of compound eye structures, which is useful for developing high-level applications. After that, a low complexity semantic segmentation scheme for compound images based on a convolutional neural network is proposed. The experimental result shows that compound images are more suitable for semantic segmentation than typical RGB images.