10 posts tagged with "computer vision"

When building computer vision systems, efficient data storage is a fundamental requirement. Whether you're capturing images for training, storing inference results for validation, or archiving sensor data for future analysis, your storage solution must be both reliable and high-performance.

Ingestion speed is especially critical. If your system can’t write data fast enough — whether it’s high-frequency frames or accompanying metadata — you risk losing valuable information or creating bottlenecks in the pipeline.

In this post, we’ll look at three common approaches to storing data in computer vision applications: a traditional file system, S3-compatible object storage, and ReductStore, a time-series-optimized blob storage. We’ll explore the strengths and limitations of each approach to help you choose the best fit for your application.

A Simple Computer Vision Application​

For demonstration purposes, we’ll use a simple computer vision (CV) application which is connected to a CV camera and runs on an edge device:

The camera driver captures images from the CV camera every second and forwards them to the model, which then detects objects and displays the results in the user interface.

Your images and results need to be stored for training and validation purposes. The customer may also wish to view images featuring anomalous objects. These requirements present the challenge of maintaining a history of blob or unstructured data.

ROS2 is widely used for building robotic systems with sensors like cameras, LiDAR, and IMUs. While it's great for communication (e.g., publishing and subscribing to topics), it lacks a built-in solution for storing large amounts of unstructured data, such as images.

Bag files are commonly used to store data in ROS2, but they aren't a good fit for long-term storage or real-time streaming. They're mainly meant for recording and replaying mission data or episodes, not for managing large volumes of unstructured data.

Addressing this challenge, this blog post will guide you through setting up ROS2 with ReductStore a high-performance storage and streaming solution optimized for unstructured, time-series data.

We will focus specifically on image data, but if you are interested in a more general overview you can read How to Store and Manage Robotic Data which covers the challenges and strategies for storing and managing robotic data in general.

For the full code example, we will be using the reduct-ros-example repository, which provides a complete implementation of the concepts discussed in this article.

A pre-trained model is a neural network that has already been trained on a large dataset to perform specific tasks, such as image classification or object detection. These models are highly valuable, allowing us to build on previous knowledge rather than starting from scratch.

However, computer vision models often require large datasets of labeled images or videos, which can quickly become challenging to manage, especially when sourced from continuous data streams. ReductStore addresses this need by providing an efficient and reliable time-series object store capable of handling large volumes of high-frequency, unstructured data such as video streams or labeled images. For practical guidance on implementing ReductStore and integrating it with Roboflow to develop high-performing computer vision models, refer to the guide: Computer Vision Made Simple with ReductStore and Roboflow.