September 2021 Summaries

Roboflow has entered into a partnership with Ultralytics, allowing it to license and integrate Ultralytics models, such as the YOLO series, into its dataset management and annotation tools, thereby becoming the official tool for Ultralytics. This collaboration enhances the integration between Roboflow and Ultralytics, enabling users to efficiently bootstrap projects with public datasets, train custom models, and iteratively improve performance by incorporating real-world data. The partnership aims to simplify the computer vision process by minimizing the complexities of machine learning and infrastructure, making it more accessible for developers to focus on domain-specific problems. Users can start using YOLO models with a free Roboflow account, and the commercial licensing options allow for deployment within business environments without the risk of violating the AGPL-3.0 license. Additionally, Roboflow is committed to supporting the open-source community by contributing financially to the Ultralytics project and encouraging innovative projects that lower barriers to accessing computer vision technology.

Jacob Solawetz's blog post explores how the OAK-D-Lite device can be used to develop a portable card-counting tool for blackjack, demonstrating the potential of computer vision applications in real-world scenarios like casinos. The post details the process of deploying custom vision models using the OAK-D-Lite, including sourcing a playing card dataset from Roboflow Universe and training a model to recognize card numbers and suits. Users can begin testing their models via hosted APIs even before receiving their device, and Roboflow facilitates the conversion of models for deployment on the OAK-D-Lite. The blog also highlights the ease of setting up an inference server and the excitement of developing card-counting logic to strategize against the house, showcasing a practical and engaging use of computer vision technology.

Roboflow's approach to MLOps emphasizes treating computer vision models as microservices to enhance separation of concerns, scalability, and speed of iteration. By deploying models as microservices, developers can isolate specific dependencies and hardware requirements from the main application, allowing for a more flexible and efficient system architecture. This approach is particularly beneficial in handling bursty usage patterns, such as monitoring security camera feeds, as it allows for more cost-effective and scalable deployment options. Additionally, microservices enable faster updates and iteration cycles for models without the need to redeploy entire applications, facilitating independent development timelines for different team members. Roboflow supports this approach through a standardized inference API, allowing models to be tested and deployed across various platforms, and simplifying updates by merely adjusting configuration files or conducting staged rollouts.

PaddlePaddle, an open-source machine learning framework developed by Baidu, is similar to other frameworks like PyTorch and TensorFlow, offering tools necessary for building deep learning models. It gained attention in the computer vision community with the introduction of PP-YOLO, a model in the YOLO family, which showcased significant advancements over its predecessors such as YOLOv4 and YOLov5. This framework is particularly noted for its parallel distributed deep learning capabilities, enhancing its utility in various machine learning applications.

Roboflow announced a $20 million Series A funding round led by Craft Ventures, with participation from notable investors, to further its mission of democratizing computer vision technology. This funding will support Roboflow's efforts to integrate computer vision into various industries by providing developers with an end-to-end platform to build high-quality models efficiently. The company emphasizes the transformative potential of computer vision, comparing its impact to that of personal computers and smartphones, and highlights its use in diverse applications like cancer research, wildlife protection, and retail improvement. With over 50,000 developers utilizing its platform, Roboflow aims to simplify the model development process, allowing teams to focus on domain-specific challenges rather than infrastructure. The company is also actively hiring to expand its team, drawing parallels to the growth trajectory of Stripe in its early days.

TensorRT is a high-performance inference acceleration library developed by NVIDIA, designed to optimize machine learning models for NVIDIA GPUs, and is considered one of the fastest ways to run models currently. Users typically start with frameworks like PyTorch or TensorFlow and convert models to TensorRT for deployment, with tools like Roboflow simplifying this transition. The installation of TensorRT involves setting up NVIDIA GPU drivers, Cuda, and TensorRT itself, preferably on a Linux base such as Ubuntu. While TensorRT is focused on GPU acceleration, alternative frameworks like OpenVINO and ONNX are recommended for CPU optimization. Additionally, TensorRT can enhance inference speeds on NVIDIA Jetson devices, with newer Jetson Jetpack distributions potentially including pre-installed TensorRT.

Jacob Solawetz's guide on the Roboflow blog provides a comprehensive tutorial on implementing object tracking using custom object detection models, specifically highlighting the ease of integrating zero-shot features to simplify the process. The guide emphasizes the importance of object tracking for applications like counting distinct objects in video streams and explains how to train an object detection model using tools such as a custom YOLOv5 model or Roboflow's training solution. It details the steps to clone the zero-shot object tracking repository, set up dependencies, and process video frames using the clip_object_tracker.py script with either YOLOv5 or the Roboflow Inference API. The tutorial also offers practical examples, including tracking cars in a video, and underscores the versatility of object tracking in various fields, encouraging users to explore its potential for diverse applications.

In September 2021, Roboflow introduced several enhancements to their product suite, notably launching Roboflow Universe, a platform for sharing computer vision projects, and offering advanced features for free to users who publicly share their data. The company unveiled new plans and pricing, improved team collaboration tools, and enabled video uploading and annotation across all plans. Major updates included the general availability of Roboflow Train, the release of a Zero Shot Object Tracking open-source repository, and the addition of a GPU inference-server for enterprise clients. The month also saw the launch of Label Assist, a refreshed logo, and team expansion with new hires. Roboflow's community engagement included volunteering, hosting an on-site team event, and participating in media outlets like Rackspace and Hacker News.