Search for: All records

Despite remarkable improvements in speed and accuracy, convolutional neural networks (CNNs) still typically operate as monolithic entities at inference time. This poses a challenge for resource-constrained practical applications, where both computational budgets and performance needs can vary with the situation. To address these constraints, we propose the Any-Width Network (AWN), an adjustable-width CNN architecture and associated training routine that allow for fine-grained control over speed and accuracy during inference. Our key innovation is the use of lower-triangular weight matrices which explicitly address width-varying batch statistics while being naturally suited for multi-width operations. We also show that this design facilitates an efficient training routine based on random width sampling. We empirically demonstrate that our proposed AWNs compare favorably to existing methods while providing maximally granular control during inference. 

Autonomous vehicles often employ computer-vision (CV) algorithms that track the movements of pedestrians and other vehicles to maintain safe distances from them. These algorithms are usually expressed as real-time processing graphs that have cycles due to back edges that provide history information. If immediate back history is required, then such a cycle must execute sequentially. Due to this requirement, any graph that contains a cycle with utilization exceeding 1.0 is categorically unschedulable, i.e., bounded graph response times cannot be guaranteed. Unfortunately, such cycles can occur in practice, particularly if conservative execution-time assumptions are made, as befits a safety-critical system. This dilemma can be obviated by allowing older back history, which enables parallelism in cycle execution at the expense of possibly affecting the accuracy of tracking. However, the efficacy of this solution hinges on the resulting history-vs.-accuracy trade-off that it exposes. In this paper, this trade-off is explored in depth through an experimental study conducted using the open-source CARLA autonomous driving simulator. Somewhat surprisingly, easing away from always requiring immediate back history proved to have only a marginal impact on accuracy in this study. 

Vision-based perception systems are crucial for profitable autonomous-driving vehicle products. High accuracy in such perception systems is being enabled by rapidly evolving convolution neural networks (CNNs). To achieve a better understanding of its surrounding environment, a vehicle must be provided with full coverage via multiple cameras. However, when processing multiple video streams, existing CNN frameworks often fail to provide enough inference performance, particularly on embedded hardware constrained by size, weight, and power limits. This paper presents the results of an industrial case study that was conducted to re-think the design of CNN software to better utilize available hardware resources. In this study, techniques such as parallelism, pipelining, and the merging of per-camera images into a single composite image were considered in the context of a Drive PX2 embedded hardware platform. The study identifies a combination of techniques that can be applied to increase throughput (number of simultaneous camera streams) without significantly increasing per-frame latency (camera to CNN output) or reducing per-stream accuracy.