Abstract:
Convolutional Neural Networks (CNNs) show tremendous performance in many Computer Vision (CV) tasks like image segmentation crucial to autonomous driving. However, they a...Show MoreMetadata
Abstract:
Convolutional Neural Networks (CNNs) show tremendous performance in many Computer Vision (CV) tasks like image segmentation crucial to autonomous driving. However, they are computationally demanding and usually not robust to image corruptions like weather influences. In this paper, we introduce our mixed-precision inference method to overcome these two challenges. Therefore, we enable mixed-precision CNN execution on modern embedded system on chips (SoC) that feature a DNN accelerator and a reconfigurable fabric. In case of a weather change, we can quickly adjust the inference precision to maintain model accuracy, while benefitting from fewer off-chip memory accesses compared to full precision. Therefore, we identify optimal quantization schemes for different weather conditions that maximize model accuracy and minimize data offloading. To enable mixed-precision inference, we present our dynamic number conversion architecture for data going back and forth to the off-chip memory, hosted on the reconfigurable tile of the SoC. Using a DeepLabV3+ model with a Resnet-101 backbone for image segmentation, for example, our evaluation yields 60% less off-chip movements under clear weather conditions. Applying rain, fog, and brightness to the input of various models, we can report an up to 26%, 23.8% and 45% reduction in data transactions, respectively, while maintaining the baseline model accuracy. We finally demonstrate that our architecture does not impact the throughput of the CNN inference and consumes very few resources of the reconfigurable fabric.
Date of Conference: 05-08 September 2023
Date Added to IEEE Xplore: 22 September 2023
ISBN Information: