Deep image compression algorithms based on Generative Adversarial Networks (GANs) are a promising direction to address the strict communication bandwidth limitations commonly encountered in IoT sensor networks (e.g. Low Power Wide Area Networks). However, current methods do not consider that the sensor nodes, which perform the image encoding, usually only offer very limited computation and memory capabilities, e.g. a resource-constrained tiny device such as a micro-controller. In this paper, we propose the first tiny generative image compression method specifically designed for image compression on micro-controllers. We base our encoder on the well-known MobileNetV2 network architecture, while keeping the decoder side fixed. To cope with the resulting asymmetric design of the compression pipeline, we investigate the impact of different training strategies (end-to-end, knowledge distillation) and integer quantization techniques (post-training, quantization-aware training) on the GAN-training stability. On the Cityscapes dataset, we achieve a compression performance that is very close to the state-of-the-art, while requiring 99% less SRAM size, 97% smaller flash storage and 87% less multiply-add operations. Our findings suggest that tiny generative image compression is particularly well suited for application-specific domains.

Many geometric invariants have been reported in the literature. For two specific differential scale invariants we discuss the parameters involved and analyze the accuracy of their computation on real camera data. Our experimental data show significant errors in the observed values. Those errors seem to be inherent in the image formation and scaling process. They indicate the limitations of local differential computations and, more generally, the problems of synthetically scaled images as models for zoomed images.

This paper presents an invariant under scaling and linear brightness change. The invariant is based on differentials and therefore is a local feature. Rotationally invariant 2-d differential Gaussian operators up to third order are proposed for the implementation of the invariant. The performance is analyzed by simulating a camera zoom-out.

The design and implementation of INTEGRA is described. INTEGRA is a range image understanding system that attempts to exploit the synergy between the various stages in the image understanding process i.e. segmentation, feature extraction, recognition and localization. Two prominent features of INTEGRA are: a synergetic combination of edge- and surface-based segmentation processes that results in more accurate segmentation than would have been possible with either of them alone; and the ability to correct errors made during segmentation in the matching and localization stages. At present, INTEGRA is capable of recognition and localization of polyhedral objects and objects that can be treated as piecewise combinations of curved surfaces of quadratic order, typically spherical, ellipsoidal, cylindrical and conical surfaces.>

Segmentation, feature extraction, recognition and localization are the four stages in range image understanding. Conventional approaches to range image understanding have treated these stages in isolation with a largely bottom-up flow of control and data through these various stages. Strictly bottom-up approaches have proved to be fragile in the face of errors in segmentation due to noise and limitations on sensor resolution and accuracy. Synergetic interaction of these various stages is essential for an image understanding system to exhibit robust behavior. This paper describes the design and implementation of INTEGRA, a range image understanding system that attempts to exploit the synergy between the various stages in the image understanding process. The salient features of INTEGRA are: (i) A synergetic combination of edge- and surface-based segmentation processes that results in more accurate segmentation than would have been possible with either of them alone and (ii) the ability to correct errors made during segmentation in the matching and localization stages. INTEGRA at this time, is limited to recognition and localization of polyhedral objects and is in the process of being enhanced to handle objects with curved surfaces of quadratic order such as spherical, ellipsoidal, cylindrical, and conical surfaces. Experimental results on real range images containing single and multiple polyhedral objects are presented. Future enhancements to INTEGRA are discussed.