Taming the Memory Demand Complexity of Adaptive Vision Algorithms

With the demand for utilizing Adaptive Vision Algorithms (AVAs) in embedded devices, serious challenges have been introduced to vision architects. AVAs may produce huge model data traffic while continuously training the internal model of the stream. This traffic dwarfs the streaming data traffic (e.g. image frames), and consequently dominates bandwidth and power requirements posing great challenges to a low-power embedded implementation. In result, current approaches either ignore targeting AVAs, or are limited to low resolutions due to not handling the traffics separately. This paper proposes a systematic approach to tackle the architectural complexity of AVAs. The main focus of this paper is to manage the huge model data updating traffic of AVAs by proposing a shift from compressing streaming data to compressing the model data. The compression of model data results in significant reduction of memory accesses leading to a pronounced gain in power and performance. This paper also explores the effect of different class of compression algorithms (lossy and lossless) on both bandwidth reduction and result quality of AVAs. For the purpose of exploration this paper focuses on example of Mixture-of-Gaussians (MoG) background subtraction. The results demonstrate that a customized lossless algorithm can maintain the quality while reducing the bandwidth demand facilitating efficient embedded realization of AVAs. In our experiments we achieved the total bandwidth saving of about 69% by applying the Most Significant Bits Selection and BZIP as the first and second level model data compression schemes respectively, with only about 15% quality loss according to the Multi-Scale Structural Similarity (MS-SSIM) metric. The bandwidth saving would be increased to 75% by using a custom compressor.