September 2022 Summaries

Computer vision, a branch of artificial intelligence, involves technologies like object detection, image classification, and keypoint detection, each serving unique functions in analyzing images and videos. Object detection identifies and locates objects within a digital image or video, often utilizing algorithms like convolutional neural networks (CNNs) or YOLO, and is widely applied in agriculture, security, and medical fields. Image classification categorizes images by analyzing pixel patterns, with models ranging from unsupervised to AI-based deep learning systems, and finds applications in areas like medical imaging and satellite imagery. Keypoint detection, meanwhile, identifies specific spatial points in images, aiding in tasks like human pose estimation and facial recognition, and is known for its capability to extract 3D features. These technologies are crucial in various industries, helping in tasks from crop counting to safety inspections, and are supported by numerous open-source datasets available for innovation and development.

Supervised and unsupervised learning are fundamental concepts in machine learning that differ primarily in the use of labeled data. Supervised learning involves training models on labeled datasets, where each input data point is paired with an output label, allowing the model to learn patterns and make predictions, such as classifying images or predicting house prices. This method is advantageous for its accuracy due to the presence of labels but can be costly and time-consuming due to the need for extensive labeled data. Common supervised learning algorithms include classification and regression methods. In contrast, unsupervised learning works with unlabeled data to identify patterns or groupings, utilizing techniques like clustering, dimensionality reduction, and association rules. Although it offers the benefit of not requiring labeled data, which can save time and resources, unsupervised learning often requires human intervention to interpret the results and might be less precise. Both approaches have their own set of applications, such as self-driving cars and spam detection for supervised learning, and data labeling and anomaly detection for unsupervised learning, each playing a critical role in advancing various business and technological fields.

The blog post by Tyler Odenthal explores using computer vision to measure distances in images and videos, focusing on the application of bounding boxes and pixel-to-real-world measurement conversions. It explains key concepts such as object detection, instance segmentation, and the use of bounding boxes to determine distances between detected objects, utilizing tools like JSON and OpenCV for visualization. The post discusses converting pixel distances into practical units like inches by establishing a pixel-to-inch ratio, which is calculated using known dimensions of reference objects, such as cars or a soda can, allowing for accurate distance estimation. It also touches on challenges such as maintaining tight bounding boxes and the resolution's impact on measurement accuracy, offering suggestions for overcoming these issues, like using multiple reference objects or incorporating additional sensors like LiDAR. The article provides insights into practical applications of this method in various fields and invites readers to explore and innovate further in their own projects.

In September 2022, a blog post by Paul Guerrie and Trevor Lynn on the Roboflow Blog detailed the process of training a YOLOv5 model for instance segmentation using a custom dataset. The post explains how YOLOv5, originally known for object detection, has expanded to support instance segmentation tasks and provides a step-by-step guide for setting up the environment, preparing a dataset, and training the model. The authors emphasize the utility of Roboflow Annotate for data labeling and augmentation, making it easier to train models by providing a comprehensive suite of tools including preprocessing options and the ability to use polygon annotations for instance segmentation. Additionally, the blog discusses using pre-trained weights to accelerate training and improve results, as well as testing and validating the custom-trained model. The authors also highlight the potential for further model improvement through active learning and deploying models to the edge using Roboflow Inference, a solution that supports deployment on various devices. The blog post serves as a comprehensive guide for those interested in leveraging YOLOv5's instance segmentation capabilities for practical applications.

Brad Dwyer's post explains how to utilize computer vision with drones to detect solar panels in aerial imagery and map their precise GPS locations through a process called georeferencing. The procedure involves training a computer vision model to identify objects in drone footage, then combining this data with drone flight logs to convert pixel coordinates into geographical coordinates. The article highlights the use of Roboflow, an end-to-end computer vision platform, to either find or train a custom model for object detection, emphasizing the importance of georeferencing to associate image data with real-world positions. It also discusses the technical aspects of converting video data into GPS coordinates, including the need for accurate drone telemetry and the application of trigonometry to determine object locations. The process is geared towards improving the accuracy and utility of drone-captured data, allowing for practical applications such as mapping detected objects on a webpage using mapping APIs. Dwyer also provides insights into tuning the model's detection confidence and handling duplicate or false predictions to enhance the reliability of the output. The post invites readers to explore further by experimenting with pre-trained models or creating custom ones, with supporting resources and tools available on GitHub.

Roboflow's July and August 2022 changelog highlights a range of updates and enhancements across its platform, including new features for organizing and annotating images, such as an Assign tab, polygon annotations, and enhanced project card UI. The platform also introduced improvements for training and deploying models, adding support for Raspberry Pi and Semantic Segmentation as a project type. Roboflow Universe received updates like new aggregate pages and enhanced project landing pages, while the team celebrated milestones such as hiring their first international member and hosting team meet-ups. The changelog also features user projects that demonstrate the application of Roboflow's tools in diverse areas such as food allergen detection, railway fault detection, and breast cancer detection, showcasing the platform's versatility in computer vision tasks.

Roboflow, an end-to-end computer vision platform, is now available through the AWS Marketplace, facilitating easier access for AWS customers who make up a significant portion of its user base, including over half of the Fortune 100. This integration allows customers to benefit from consolidated billing and streamlined procurement processes, leveraging their existing AWS Enterprise agreements and enhancing cost management through the AWS Cost Explorer. The listing offers flexible purchasing options, including the Roboflow Growth license and customizable enterprise-level solutions, aligning with Roboflow's mission to democratize computer vision by utilizing AWS's infrastructure to deploy its model pipeline broadly. This collaboration aims to break down barriers in machine learning adoption, making advanced visual AI more accessible to developers.

Jacob Witt describes how he developed a computer vision system to monitor public tennis court availability using a webcam and a Python script, despite having limited machine learning experience. He utilized Roboflow for data labeling and model training, creating a dataset with images captured over two days to account for varying conditions. The system identifies the presence of tennis players using labeled data and sends real-time notifications via Twilio when courts become available. Witt's project showcases an accessible approach to applying computer vision technology for personal use, demonstrating the potential for future enhancements like automating regular checks or expanding the system to cover multiple locations.

Every summer, North Carolina State University's College of Engineering hosts week-long camps for high school students to experience engineering, particularly focusing on computer vision and machine learning. Under the guidance of Dr. Edgar Lobaton, Ph.D. candidates Mohamed Traore, Jeremy Park, and Sanjana Banerjee developed activities that walked students through the computer vision pipeline using Roboflow, from data collection to model prediction. The students used Bluetooth sensors to collect acceleration data during standing, walking, and running, which was then converted into spectrograms to serve as training data for machine learning models. Participants explored Roboflow Universe, learned about image classification, and trained their own activity recognition models with high performance. The initiative aimed to inspire students to create future computer vision applications, with one student expressing interest in developing a skin disease detection tool. The program was supported by NSF grants and emphasized the potential of computer vision in solving global challenges.