Towards Communication-Efficient Distributed Background Subtraction

Road traffic monitoring is one of the essential components in data analysis for urban air pollution prevention. In road traffic monitoring, background subtraction is a critical approach where moving objects are extracted via facilitating motion information of interest to static surroundings, known as backgrounds. To work with various contextual dynamics of nature scenes, supervised models of background subtraction aim to solve a gradient-based optimization problem on multi-modal sequences of videos by training a convolutional neural network. As video datasets are scaling up, distributing the model learning on multiple processing elements is a pivotal technique to leverage the computational power among various devices. However, one of major challenges in distributed machine learning is communication overhead. This paper introduces a new communication-efficient distributed framework for background subtraction (CEDFrame), alleviating the communication overhead in distributed training with video data. The new framework utilizes event-triggered communication on a ring topology among workers and the Partitioned Globally Address Space (PGAS) paradigm for asynchronous computation. Through the new framework, we investigate how training a background subtraction tolerates the tradeoffs between communication avoidance and accuracy in model learning. The experimental results on NVIDIA DGX-2 using the CDnet-2014 dataset show that the new framework can reduce the communication overhead by at least 94.71% while having a negligible decrement in testing accuracy (at most 2.68%).