Towards closing the energy gap between HOG and CNN features for embedded vision

Abstract

Computer vision enables a wide range of applications in robotics/drones, self-driving cars, smart Internet of Things, and portable/wearable electronics. For many of these applications, local embedded processing is preferred due to privacy and/or latency concerns. Accordingly, energy-efficient embedded vision hardware delivering real-time and robust performance is crucial. While deep learning is gaining popularity in several computer vision algorithms, a significant energy consumption difference exists compared to traditional hand-crafted approaches. In this paper, we provide an in-depth analysis of the computation, energy and accuracy trade-offs between learned features such as deep Convolutional Neural Networks (CNN) and hand-crafted features such as Histogram of Oriented Gradients (HOG). This analysis is supported by measurements from two chips that implement these algorithms. Our goal is to understand the source of the energy discrepancy between the two approaches and to provide insight about the potential areas where CNNs can be improved and eventually approach the energy-efficiency of HOG while maintaining its outstanding performance accuracy.

ML‐morph: A fast, accurate and general approach for automated detection and landmarking of biological structures in images

“…In our view, these limitations can be easily overcome using one of the following: (a) training of multiple object detectors, one for each position or (b) through the use of CNN‐based detectors, such as Ren et al (). While CNN‐based detectors are insensitive to changes in orientation, it is worth pointing out that such algorithms require a larger training dataset (more images), higher hardware specifications, are less portable (file size), and require more fine‐tuning (Suleiman et al, ).…”

“…They also require large training samples (one to two orders of magnitude higher than the one proposed here), more fine‐tuning of training parameters, specialized hardware (e.g. graphics processing unit, GPU), and usually perform at lower image processing rates (images per second) than simpler models (Suleiman, Chen, Emer, & Sze, ).…”

ML-morph: A Fast, Accurate and General Approach for Automated Detection and Landmarking of Biological Structures in Images

“…While the current state of the art in object detection relies on convolutional neural networks (CNN, e.g., Ren et al, 2015), these can be overpowered for standard biological applications. They also require large training samples (one to two orders of magnitude higher than the one proposed here), more fine-tuning of training parameters, specialized hardware (e.g., GPU), and usually perform at lower image processing rates (images per second) than simpler models (Suleiman et al, 2017).…”

“…In our view, these limitations can be overcome using one of the following: (1) training of multiple object detectors, one for each position or (2) through the use of CNN-based detectors, such as Ren et al (2015). While CNN based detectors are insensitive to changes in orientation, it is worth pointing out that such algorithms require a larger training dataset (more images), higher hardware specifications, are less portable (file size), and require more fine tuning (Suleiman et al, 2017).…”

Efficient ConvNet-based Object Detection for Unmanned Aerial Vehicles by Selective Tile Processing

“…is is particularly important in certain cases such as cameras on-board UAVs where he altitudes that UAVs y/hover, and the resulting resolution of objects in the image, make the detection of small objects even more challenging. However, increasing the CNN receptive eld makes the learning process di cult since the problem dimensionality is increased and at the same time the computational cost also increases exponentially [9]. To tackle this issue it is necessary to reduce the data that needs to be processed from the original input image.…”