September 2020 Summaries

The VGG Image Annotator is a free, open-source tool developed by researchers at Oxford University for annotating images and videos to create datasets for computer vision models. This lightweight tool can be run entirely in a web browser, allowing users to draw bounding boxes or polygons around objects and facilitating collaboration in large-scale labeling projects. The tutorial provides a step-by-step guide on how to start a project, load images, label them, assign class labels, and export the annotations as a CSV file. Additionally, it highlights Roboflow's capabilities for converting VGG annotations into over 30 different computer vision formats, enhancing the ease and efficiency of managing datasets.

Amitabha Banerjee successfully trained his Anki Vector robot to detect other robots using YOLOv5 and Roboflow, a project with potential real-world implications for robotic collaboration. Banerjee collected and annotated 303 images of robots in various conditions, processed them with Roboflow for training, and used Google Colab and a T4 GPU to train his model over 50 epochs. The results showed promising performance, with the model performing well in different video backgrounds. Banerjee has shared his dataset for public use and hopes to see further improvements and applications from the community, while also anticipating the open-source release of Vector's firmware to deploy the model directly onto the robot.

CreateML is a machine learning tool by Apple that allows users to develop models for images, sound, text, and more within the Apple ecosystem. It offers a no-code environment for training models, particularly useful in computer vision tasks like image classification and object detection. While CreateML simplifies the model training process, users must convert data into Apple's proprietary JSON format, often requiring additional steps unless they use a tool like Roboflow for conversion. To build a computer vision model, users collect and label data, configure an object detection model in CreateML, and train it, monitoring metrics like validation loss and Intersection Over Union to evaluate performance. The platform allows for iterative improvements by retraining models with additional or augmented data, aiming to enhance the model's generalization and accuracy. Once trained, models can be integrated into applications, exemplified by deploying them in live camera feeds on iPhones.

An annotation group in the context of image datasets refers to the overarching category that includes all classes within a dataset, serving the purpose of organizing and labeling images in various ways depending on the training goals for a model. This concept is particularly useful when the same image needs to be labeled differently for different models, such as distinguishing between different board games and their respective pieces in an augmented reality application. Roboflow utilizes annotation groups to enhance dataset management by allowing images to have multiple annotations across various datasets while maintaining image count efficiency and enabling the correction of annotations across these datasets. Choosing an annotation group involves selecting a specific name that best represents all classes in a dataset, which aids in maintaining a scalable ontology as the dataset library expands. While generic terms like "object" or "thing" could be used, they might not be ideal for scalability in larger datasets, and selecting a more specific term is recommended to avoid future complications.

In a recent Fireside chat, Roboflow machine learning engineer Jacob Solawetz shared his professional journey from financial trading to artificial intelligence, focusing on the transition from natural language processing to computer vision and its future prospects. He highlighted key use cases at Roboflow, offered insights on building successful models, and addressed audience questions, particularly on Roboflow's support for transfer learning. Transfer learning, which leverages previously learned model parameters for new tasks, benefits from Roboflow's versioned datasets and the Roboflow Model Library that provides access to pre-trained weights. The success of this approach is contingent on the similarity between tasks. The session invited further inquiries and encouraged participation in future events through newsletter subscriptions.

The blog post by Jacob Solawetz provides a detailed guide on deploying the YOLOv5s object detection model to an NVIDIA Jetson Xavier NX at 30 FPS, utilizing Roboflow Deploy to simplify the process. It emphasizes the importance of edge AI for real-time computer vision on affordable devices and outlines steps such as training a custom YOLOv5 model, setting up the Jetson NX with NVIDIA Jetpack, and using PyTorch in Docker containers for deployment. It also suggests enhancing model performance by exploring TensorRT solutions and highlights Roboflow's ability to manage datasets, train models, and deploy to various platforms efficiently. The post concludes with additional resources for selecting the right camera and lens for computer vision projects.

The concept of the train, validation, and test split is crucial in machine learning to prevent model overfitting and ensure accurate evaluation in computer vision projects. The training set, typically comprising 70-80% of the data, is used to fit the model, while the validation set, about 10-20%, helps gauge its performance during training, guiding adjustments and early stopping. The test set, also around 10%, evaluates the model's final performance in a real-world scenario, ensuring it hasn't been tailored to the validation metrics. Effective data preprocessing and augmentation are essential, with augmentations applied only to the training set to enhance its size, while preprocessing standardizes data across all splits. Common pitfalls include train/test bleed, where similar images appear in different splits, and overemphasis on either training or validation/test metrics, potentially skewing evaluation outcomes. Roboflow offers tools to manage these processes, automatically handling issues like duplicates, ensuring the integrity of the train, validation, and test splits crucial for robust model deployment.

The article provides a comprehensive guide on training a custom image classification model using the Resnet34 architecture and the fastai library, leveraging pre-trained weights from the ImageNet dataset for transfer learning. It introduces the use of Roboflow for dataset management, including data preprocessing and augmentation to enhance model performance. The tutorial covers steps from data preparation, model fine-tuning, training with fastai, to evaluating model performance with test inference and visualizing a confusion matrix. Techniques such as freezing and unfreezing model layers, employing callbacks like EarlyStopping, and adjusting learning rates are discussed to optimize training. The guide emphasizes the ease of using Roboflow to manage datasets and deploy models, encouraging readers to try building computer vision models efficiently with its resources.