ESDA: A Composable Dynamic Sparse Dataflow Architecture for Efficient Event-based Vision Processing on FPGA

<p dir="ltr">ESDA is a framework for building customized DNN accelerators for event-based vision tasks. It leverages the spatial sparsity of event-based input by a novel dynamic sparse dataflow architecture. This is achieved by formulating the computation of each dataflow module as a unified token-feature computation scheme. To enhance the spatial sparsity, ESDA also constrained <a href="https://arxiv.org/abs/1706.01307" rel="nofollow" target="_blank">Submanifold Sparse Convolution</a> to build our DNN models.</p><p dir="ltr">The project mainly consists of three parts</p><ul><li>Software model training on event-based datasets with sparsity and quantization</li><li>Hardware design optimization (use synthesis optimization to search for optimal mapping)</li><li>Hardware synthesis, implementation, and evaluation</li></ul><p></p>