[Paper Review] Large-scale image analysis using docker sandboxing

Biswa Sengupta, Eduard Vazquez+5 more|arXiv (Cornell University)|Mar 7, 2017

Advanced Neural Network Applications22 references3 citations

TL;DR

This paper presents Cortexica, a scalable, Docker-based architecture for large-scale image localization and retrieval that leverages GPU acceleration and containerized microservices to enable fast, flexible, and secure multi-product search in commercial applications. The system combines deep learning for object detection with distributed inverted indexing for efficient retrieval, achieving sub-second response times across diverse image databases with high accuracy and fault tolerance.

ABSTRACT

With the advent of specialized hardware such as Graphics Processing Units (GPUs), large scale image localization, classification and retrieval have seen increased prevalence. Designing scalable software architecture that co-evolves with such specialized hardware is a challenge in the commercial setting. In this paper, we describe one such architecture ( extit{Cortexica}) that leverages scalability of GPUs and sandboxing offered by docker containers. This allows for the flexibility of mixing different computer architectures as well as computational algorithms with the security of a trusted environment. We illustrate the utility of this framework in a commercial setting i.e., searching for multiple products in an image by combining image localisation and retrieval.

Motivation & Objective

• To design a scalable, secure, and flexible software architecture for large-scale image analysis in commercial settings.
• To address the challenge of co-evolving software with specialized hardware like GPUs while maintaining algorithmic and architectural flexibility.
• To enable efficient, low-latency multi-product search by combining object localization and content-based retrieval.
• To support hot-swapping of deep learning models and alternative algorithms within a trusted, isolated environment using Docker containers.
• To ensure system robustness and performance through distributed, sharded inverted indexing and parallel query execution.

Proposed method

• Deploying a microservices architecture using Docker containers to sandbox each component, ensuring isolation, security, and ease of hot-swapping algorithms.
• Using state-of-the-art deep learning models—Faster R-CNN, SSD, and R-FCN—for object localization with GPU acceleration to achieve real-time inference.
• Implementing a multi-scale feature extraction pipeline with region proposal networks and shared convolutional backbones (e.g., VGG) for efficient detection.
• Applying color-texture descriptors (e.g., SIFT, LBP) for content-based image retrieval after localization, enabling similarity search in large databases.
• Constructing distributed, sharded inverted indexes that map visual words to image IDs, enabling horizontal scaling and fault tolerance.
• Executing retrieval queries in parallel across multiple database instances, each dedicated to a specific product category (e.g., tops, jackets), with load balancing and consistent response times.

Experimental results

Research questions

• RQ1 How can a scalable and secure software architecture be designed to co-evolve with specialized hardware like GPUs for large-scale image analysis?
• RQ2 What architectural patterns enable efficient hot-swapping of deep learning models and alternative algorithms in a production environment?
• RQ3 How can object localization and content-based retrieval be effectively combined to enable multi-product search in a single image or video stream?
• RQ4 What indexing strategy ensures low-latency, scalable, and fault-tolerant retrieval across large, heterogeneous image databases?
• RQ5 To what extent can a modular, containerized framework maintain high performance and accuracy while supporting diverse algorithms and data types?

Key findings

• The Cortexica framework achieves sub-second response times for multi-product search across 100k to 2M image databases, regardless of database size or number of detected items.
• The use of Docker sandboxing enables seamless hot-swapping of neural networks and alternative algorithms without service interruption, enhancing system flexibility.
• Splitting the retrieval database by product category (e.g., tops, jackets) significantly improves inference accuracy compared to a monolithic database.
• Distributed inverted indexing with sharding ensures fault tolerance: a single server failure affects only a small subset of words, minimizing impact on query results.
• Parallel execution of retrieval queries across category-specific databases maintains consistent response times and prevents performance degradation with increasing detection count.
• The system demonstrates high accuracy in localizing and retrieving multiple fashion items from a single image or video stream, with visual results showing semantically relevant matches.