July 2020 Summaries

Jacob Solawetz provides a comprehensive tutorial on how to train the EfficientNet convolutional neural network using a custom dataset for image classification. EfficientNet, developed by Google Brain, is renowned for its ability to efficiently scale convolutional neural networks, making it highly effective for classification tasks. The tutorial, hosted on Google Colab, focuses on classifying rock, paper, scissors hand gestures but is adaptable to any classification type with proper dataset supervision. Using the Keras framework, the tutorial guides users through setting up their training environment, selecting the appropriate EfficientNet model variant, and preparing their dataset using tools like Roboflow for preprocessing and augmentation. It emphasizes the importance of customizing the model architecture to fit the number of classes in the dataset and provides insights on monitoring training progress through metrics like accuracy and loss. The tutorial also covers fine-tuning, inference testing, and saving model weights for future use, ultimately equipping readers with the skills to leverage EfficientNet for various image classification applications.

Jacob Solawetz's blog post provides a comprehensive analysis of three major cloud providers' no-code tools for training custom object detection models: Amazon Rekognition Custom Labels, Azure Custom Vision, and Google Cloud AutoML Vision. The post evaluates these platforms based on factors like training time, costs, model performance, and inference time and costs, comparing them against the open-source YOLOv5 model. The evaluation reveals that while open-source models can perform on par with these cloud services, each platform presents unique strengths and limitations, such as proprietary APIs, ease of use, and varying cost structures. For those with small budgets seeking to explore computer vision, Amazon Rekognition is recommended, while those with larger budgets may benefit from using Roboflow to test all platforms and optimize for performance or sustainable long-term costs. The blog underscores the importance of considering project-specific requirements, such as inference speed and budget, and suggests Roboflow Train as an alternative solution for deploying state-of-the-art models.

Jacob Solawetz's blog post discusses the process of annotating images for computer vision models using Microsoft's Visual Object Tagging Tool (VoTT), a free, open-source solution. The post outlines the steps for gathering a dataset, installing VoTT, and creating an annotation project, emphasizing the importance of narrowing the dataset's domain for better modeling results. Solawetz provides guidance on using VoTT's shortcuts and best practices for labeling images, such as maintaining tight bounding boxes around objects and resolving any ambiguities in annotations. Once labeled, datasets can be exported in formats like Pascal VOC, which can then be uploaded to Roboflow for further management, conversion to other formats, and data augmentation. The article highlights VoTT's advantages over Roboflow while also showcasing Roboflow's additional features, including dataset management, one-click model training, and deployment options.

Result! Data, a Netherlands-based consultancy, has developed an application called Spobber to automate the detection of road signs, specifically hectometre signs on Dutch highways and roads, using computer vision. By leveraging various object detection models like YOLOv3, EfficientDet, YOLOv4, and YOLOv5, the app aims to identify these signs accurately and quickly, even at high speeds, such as 100 kilometers per hour. The development involved collecting and annotating a dataset of images, performing data augmentation to simulate various conditions, and using a split of training, validation, and test sets to prevent overfitting. The YOLOv5 model emerged as the most effective due to its speed and efficiency, though the researchers emphasize the importance of diverse training data to ensure models can handle a wide range of scenarios effectively. Future plans include enabling real-time mobile detection of objects to facilitate immediate responses to changes in field conditions.

In a detailed exploration of using the YOLOv5 deep learning algorithm for retail item detection, this article illustrates how the model can efficiently identify and classify objects on grocery store shelves, thereby aiding inventory management. The process involves leveraging the SKU110k dataset and the Roboflow platform for image annotation and data management to train a customized YOLOv5 model, which is noted for its speed and reduced model size compared to previous versions. The model was trained on a selection of images using Google Colab's GPU resources, demonstrating significant performance with metrics such as a mean Average Precision of 0.7 and a recall rate of 0.8. The article emphasizes the model's potential for real-time applications in smart retail environments, where it can manage store inventories or facilitate automatic checkout processes. Despite the computational resources required for training, YOLOv5's efficiency in real-time object detection and its adaptability make it a compelling choice for embedded systems in retail and other industries.

The blog post by Jacob Solawetz offers a detailed tutorial on training custom object detection models using the TensorFlow 2 Object Detection API, leveraging tools like Google Colab for free GPU resources and the Roboflow platform for data management and augmentation. Highlighting the ease of training state-of-the-art models such as EfficientDet, the post guides users through steps including dataset preparation, model configuration, and training execution. It also covers exporting the trained models for inference and provides insights into using TensorBoard for monitoring training progress. The API supports flexibility in switching between different computer vision techniques, making it suitable for detecting a wide range of custom objects. The tutorial aims to equip users with the knowledge to harness deep learning technologies for object detection effectively.

The TensorFlow Object Detection API has been updated to TensorFlow 2.0, bringing significant advancements for computer vision developers. The integration of Roboflow's model library with the TensorFlow 2.0 Object Detection API allows for swift training of custom models, requiring minimal code modification for data import. The API provides a comprehensive model zoo, including the EfficientDet models known for their state-of-the-art performance on the COCO dataset. TensorFlow 2.0 introduces eager execution, simplifying the development process by allowing operations to be defined and executed dynamically. Additionally, the API supports training distribution across multiple GPU or TPU resources, enhancing scalability for large datasets. This update promises improved performance and efficiency for object detection tasks, encouraging developers to utilize the new features and capabilities available in TensorFlow 2.0.

Anchor boxes are essential components in object detection models, enabling them to accurately predict and localize multiple objects within an image by serving as a starting point for bounding box predictions. These models, like EfficientDet and YOLO, use anchor boxes to hypothesize locations for objects, which are then refined through regression and classification processes. The model's performance can be significantly enhanced by carefully tuning anchor boxes, especially when dealing with irregularly shaped objects. For instance, YOLOv5 can auto-adjust anchor box distributions based on the training set, which is beneficial for datasets with objects that deviate from standard shapes found in common datasets like COCO. Customizing anchor boxes, particularly for datasets with unique object shapes such as tall or wide objects, can help improve the model's accuracy in localizing true objects in the image, making anchor box configuration a crucial step in developing robust object detection models.

Jacob Solawetz's blog post provides a detailed guide on converting Supervise.ly annotations to the YOLO Darknet format, enabling users to utilize Supervise.ly data for custom computer vision models outside of its native platform. The process involves downloading annotated datasets from Supervise.ly, which are initially in a custom JSON format, and then converting these annotations into the simpler YOLO Darknet format using Roboflow. This conversion includes creating a class map in a .labels file and specifying bounding box coordinates in individual .txt files. The post also highlights the flexibility of the YOLO Darknet format for training object detectors and suggests using Roboflow's tools for quick and efficient conversion to other annotation formats such as TFRecord, Pascal VOC, and COCO JSON, among others. Additionally, the guide encourages users to leverage their newly formatted data by training a model, particularly recommending a tutorial on training a YOLOv5 object detector.

Roboflow introduces Class Label Remapping and Omission as a feature that allows users to efficiently manage class labels in their annotated object detection datasets. This tool enables users to address and rectify computer vision labeling errors, optimize model creation by simplifying class scopes, and handle class imbalances by omitting underrepresented or overlapping classes. The process involves navigating the Roboflow platform to modify class labels through preprocessing steps, enabling users to rename, remap, and omit labels as needed. This feature not only helps correct errors from labeling jobs but also aids in enhancing model performance by allowing users to experiment with different class configurations and reduce complexities. By merging datasets and adjusting class labels, users can increase training data efficiency and improve the robustness of their models, especially when dealing with sparse datasets or overlapping annotations.

Roboflow distinguishes itself by complementing outsourced labeling services such as Scale, LabelBox, and Amazon SageMaker Ground Truth, rather than competing directly with them. It facilitates easy integration by supporting importation of annotations from these services' APIs and allows Pro tier users to import SageMaker Ground Truth manifest files directly from an S3 bucket. Roboflow provides tools for dataset visualization and quality control, making it an effective platform for testing multiple labeling providers and avoiding vendor lock-in. The platform seamlessly integrates with popular annotation tools like CVAT and LabelImg, streamlining the process of incorporating labeled images into various training pipelines. For those interested in learning more about its labeling integrations, Roboflow offers demonstrations through their sales team.

YOLOv4-tiny, a streamlined version of the YOLOv4 object detection model, offers significant speed advantages in both training and inference, making it especially useful for those with limited computational resources. While it trades off some accuracy for this increased speed, it remains effective for many tasks, particularly in cases involving smaller custom datasets. The article provides a detailed guide on setting up a training environment using Google Colab, configuring the model for custom datasets, and executing the training process, emphasizing the ease of use provided by platforms such as Roboflow. It highlights the architectural differences between YOLOv4-tiny and its larger counterpart, noting the reduced number of layers and anchor boxes. The guide also touches on the importance of adjusting training parameters based on dataset characteristics and monitors the mean average precision to avoid overfitting. After training, the model can quickly infer on test images, and the weights can be exported for further applications without the need for retraining.

The Maasai, an indigenous ethnic group in Kenya and northern Tanzania, have faced significant land displacement due to British colonialism and recent land grabs by the Kenyan and Tanzanian governments under the guise of conservation and tourism. To address this, students from General Assembly are utilizing satellite data from Google Earth Engine and computer vision models to monitor Maasai villages, track their movements, and assess changes in land usage. This initiative aims to create an automated system to identify village structures and provide a form of surveillance against unjust land claims, with the project leveraging Roboflow's preprocessing techniques to enhance model performance. The team hopes to present their findings to the Maasai community and see their model used as an AI Human Rights Watchdog to protect the rights of the Maasai people.