Abstract:This paper presents a scalable deep learning model called Agile Temporal Convolutional Network (ATCN) for high-accurate fast classification and time series prediction in resource-constrained embedded systems. ATCN is primarily designed for mobile embedded systems with performance and memory constraints such as wearable biomedical devices and real-time reliability monitoring systems. It makes fundamental improvements over the mainstream temporal convolutional neural networks, including the incorporation of separable depth-wise convolution to reduce the computational complexity of the model and residual connections as time attention machines, increase the network depth and accuracy. The result of this configurability makes the ATCN a family of compact networks with formalized hyper-parameters that allow the model architecture to be configurable and adjusted based on the application requirements. We demonstrate the capabilities of our proposed ATCN on accuracy and performance trade-off on three embedded applications, including transistor reliability monitoring, heartbeat classification of ECG signals, and digit classification. Our comparison results against state-of-the-art approaches demonstrate much lower computation and memory demand for faster processing with better prediction and classification accuracy. The source code of the ATCN model is publicly available at https://github.com/TeCSAR-UNCC/ATCN.
Abstract:Recent years have brought great advancement in 2D human pose estimation. However, bottom-up approaches that do not rely on external detectors to generate person crops, tend to have large model sizes and intense computational requirements, making them ill-suited for applications where large computation costs can be prohibitive. Lightweight approaches are exceedingly rare and often come at the price of massive accuracy loss. In this paper, we present EfficientHRNet, a family of lightweight 2D human pose estimators that unifies the high-resolution structure of state-of-the-art HigherHRNet with the highly efficient model scaling principles of EfficientNet to create high accuracy models with significantly reduced computation costs compared to other state-of-the-art approaches. In addition, we provide a formulation for jointly scaling our backbone EfficientNet below the baseline B0 and the rest of EfficientHRNet with it. Ultimately, we are able to create a family of highly accurate and efficient 2D human pose estimators that is flexible enough to provide a lightweight solution for a variety of application and device requirements. Baseline EfficientHRNet achieves a 0.4% accuracy improvement when compared to HRNet while having a 34% reduction in floating point operations. When compared to Lightweight OpenPose, a compressed network designed specifically for lightweight inference, one EfficientHRNet model outperforms it by over 10% in accuracy while reducing overall computation by 15%, and another model, while only having 2% higher accuracy than Lightweight OpenPose, is able to further reduce computational complexity by 63%. At every level, EfficientHRNet proves to be more computationally efficient than other bottom-up 2D human pose estimation approaches, while achieving highly competitive accuracy.