Efficient Embedded Deep Neural-Network-based Object Detection Via Joint Quantization and Tiling

Embedded visual AI is a growing trend in applications requiring low latency, real-time decision support, increased robustness and security. Visual object detection, a key task in visual data analytics, has enjoyed significant improvements in terms of capabilities and accuracy due to the emergence of Convolutional Neural Networks (CNNs). However, such complex paradigms require heavy computational resources that prevent their deployment on resource-constrained devices, and in particular, impose significant constraints in possible hardware accelerators geared towards such applications. In this work therefore, we investigate how a combination of techniques can lead to efficient visual AI pipelines for resource-constrained object detection. In particular we leverage an efficient search strategy based on a combination of pre-processing mechanisms, that reduce the processing demands of deep network as a counter measure for potential accuracy reduction caused by quantization. The proposed approach enables the detection of objects in higher resolution frames using quantized models, while maintaining the accuracy of full-precision CNN-based object detectors. We illustrate the impact on the accuracy and average processing time using quantization techniques and different tiling approaches on efficient object detection architectures; as a case study, we focus on Unmanned-Aerial- Vehicles (UAVs). Through the proposed methodology, hardware accelerator demands are thereby reduced, leading to both performance benefits and associated power savings.