Lambda-IVR: An Indexing Framework for Video Big Data Retrieval Using Distributed In-memory Computation in the Cloud

Feature indexing for video retrieval poses a significant hurdle for indexing due to three significant challenges. First, there are different types of features in varying nature, such as deep Convolutional Neural Network (CNN) features, handcrafted features, recognized text from the videos, and audio features, etc. Secondly, feature matching for those varying types of features requires different similarity measure methods. And thirdly, considering the Big-Data era the number of features to be indexed is enormous. To address these issues, in this paper, we present a lambda style distributed in-memory scale-out inverted-index based feature indexing framework for video retrieval, which operates as SaaS in the cloud. First, the video features are acquired, decoupled, and the visual features are encoded using an adaptation of an existing feature encoder with improvements. Secondly, the visual encoded features and the textual features are aggregated. Finally, the aggregated features are indexed and readily available for retrieval. Our framework supports incremental updates without the need to re-index the data and can serve enormous concurrent queries. Experimental results show that our framework performs reasonably well in terms of, accuracy, precision, and efficiency.