August 2020 Summaries

Cocoparks, a Paris-based startup focused on improving traffic flows in French cities, accelerated their development process significantly by using Roboflow, a platform likened to the "Google Maps" of computer vision development. The startup utilized Roboflow to streamline their workflow, allowing them to launch a minimum viable product in just 30 days. Key improvements included efficient data preparation and augmentation, which enabled the team to enhance their training datasets quickly and visualize and rectify labeling issues. Roboflow also facilitated seamless conversion between various model frameworks, saving the Cocoparks team substantial engineering time and allowing them to focus on improving model performance. By leveraging Roboflow's tools, Cocoparks achieved a 93% mean average precision (mAP) on their task, demonstrating significant model improvement and allowing them to validate their idea and attract more customers rapidly.

Fastai v2, a significant update to the popular deep learning framework, introduces improvements to its library and offers new resources, including a machine learning course and helper repositories. Created by Jeremy Howard, Rachel Thomas, and Sylvain Gugger, fastai aims to democratize artificial intelligence by making it accessible and understandable to everyone through a layered architecture that allows users to perform complex AI tasks with minimal code. The new version enhances flexibility and ease of use, boasting a consistent experience across domains such as NLP and computer vision, and is built on top of PyTorch, incorporating updates to its functionality. Key features include an infinitely customizable training loop facilitated by callbacks and GPU-accelerated image augmentations. Additionally, fastai v2 introduces helper libraries like fastcore, fastscript, and fastgpu, which provide additional functionality and ease of use for Python developers. The release is complemented by a lecture series that explores foundational machine learning concepts and demonstrates the application of the new library.

Roboflow, a company co-founded by Joseph Nelson and Brad Dwyer, aims to democratize computer vision technology, allowing any developer to integrate it into their projects without needing expertise in machine learning. The company joined Y Combinator's Summer 2020 batch, marking a significant milestone as this was the first time the accelerator program was conducted entirely remotely. The team set up a workspace in Des Moines, Iowa, utilizing technology to collaborate effectively across multiple time zones and benefiting from the extensive Y Combinator network. Participation in the program has accelerated Roboflow's development, focusing on customer needs and facilitating faster service delivery. The supportive YC community has also seen several startups adopting Roboflow's tools, further validating its mission to integrate computer vision into everyday applications.

Brad Dwyer, co-founder and CTO of Roboflow, discusses a webinar co-hosted with Mark McQuade from Onica, an AWS Premier Consulting Partner, focusing on the integration of Roboflow with AWS Rekognition Custom Labels for training and deploying custom object detection models. The video tutorial specifically demonstrates the process of creating a mask/no mask detector, providing a comprehensive guide from start to finish. Viewers are encouraged to engage with the content by subscribing to the YouTube channel for more insights into computer vision. Dwyer, who has a background in developing computer vision infrastructure and founding successful ventures like Hatchlings, emphasizes the importance of staying updated with the latest advancements in the field.

Jamie Shaffer, a data scientist based in Washington state, explores the application of deep learning, specifically object detection, to count salmon passing through fish ladders in the Pacific Northwest. This region's salmon populations are crucial for environmental assessment, and the task of counting them is traditionally labor-intensive, involving trained experts to manually review video recordings. Shaffer utilizes the YOLO v5 model for its speed and accuracy, leveraging open-source tools like Roboflow for data augmentation and model training. Despite challenges such as varying lighting and image quality, the model achieves a mean Average Precision ([email protected]) of 70%, suggesting that machine learning can effectively assist in fish counting. This approach not only streamlines the process but also provides valuable data for managing fishing seasons and dam operations, highlighting the readiness of AI for real-world environmental applications.

The blog post by Jim Su and Samrat Sahoo provides a comprehensive guide on training a custom mobile object detection model using the YOLOv4 tiny Darknet model and converting it to TensorFlow Lite for on-device inference. It details the steps required to prepare custom data, train the YOLOv4 tiny model using the Darknet framework, and then convert the trained model to TensorFlow Lite, which is suitable for on-device deployment. The tutorial emphasizes the use of Roboflow for data management, labeling, and conversion, showcasing how to deploy the model on Android devices using TensorFlow Lite, including a practical demonstration with an Expo app. The guide also highlights the advantages of using TensorFlow Lite for on-device inference and provides insights into deploying models on different platforms like iOS, making it a valuable resource for those interested in building and deploying custom object detection models.

Roboflow, a computer vision dataset management tool, successfully acquired the domain roboflow.com after it had been owned by the German fluid control company Bürkert since 2001. Initially, the domain redirected to Bürkert's homepage after they acquired a company called Robolux in 2005, which had a product named Roboflow, but the product was discontinued and its trademark expired in 2009. Despite initial doubts about obtaining the domain from such a large company, the Roboflow team decided to pursue it as their startup grew, strategically reaching out to Bürkert before announcing any funding to avoid a price hike. After navigating through company contacts, they were eventually offered the domain at a reasonable price, which they accepted, marking a significant milestone in their journey to democratize computer vision for software developers.

The article provides a comprehensive guide on how to train and export a custom TensorFlow Lite object detection model using a personal dataset, tailored for deployment on low-performance devices like mobile phones and IoT hardware. It walks through the process of preparing object detection data in TFRecord format using Roboflow, a platform that facilitates data management, labeling, and conversion. The tutorial emphasizes the use of MobileNet Single Shot Detector (v2) architecture, optimized for lightweight inference, and includes steps for configuring a training pipeline in a Colab Notebook. After training a TensorFlow model, the guide explains the conversion of the SavedModel to TensorFlow Lite using the command line converter, with instructions for deploying the model on Android, iOS, or Raspberry Pi. Additionally, Roboflow's capabilities in managing datasets, training models, and deploying them across different platforms are highlighted, offering a streamlined approach for building computer vision applications.

Joseph Nelson's blog post discusses Charles Herring's initiative to utilize computer vision to help tool brands identify their presence in YouTube videos, aiming to leverage the platform's massive video uploads for brand exposure and marketing opportunities. Charles, a software engineer, developed a model to detect when tools are featured in videos, offering insights into brand visibility and potential sponsorships for content creators. Initially using Google Cloud Platform for data management, Charles faced challenges with limited and diverse datasets. He transitioned to Roboflow Pro, which improved his workflow by enhancing dataset management and augmentation, resulting in a significant increase in the model's accuracy from a mean average precision (mAP) of 0.817 to 0.885. This efficiency allowed Charles more time to focus on his core objectives and market his services to major power tool retailers.

Transport for Cairo (TfC) is a data-driven consultancy that focuses on improving urban mobility in emerging African cities by addressing traffic congestion and rapid urbanization, particularly in the context of climate change. In a recent project in Uganda, where traffic data was scarce, TfC utilized computer vision to efficiently and accurately monitor traffic flows. This was made possible by leveraging existing closed-circuit TV infrastructure and developing custom models to recognize a variety of vehicles unique to the area, a task that pre-trained models could not accomplish. TfC used Roboflow to process labeled data, manage class imbalances, and augment datasets, which streamlined the training of their computer vision models. This approach not only saved time and resources but also enhanced the accuracy and reliability of traffic data collection, allowing TfC to optimize transportation systems and develop sustainable mobility solutions for the future.

Baidu's PP-YOLO introduces a significant advancement in object detection by building on YOLOv3 and leveraging the PaddlePaddle deep learning framework, offering faster inference speeds and improved accuracy over YOLOv4. PP-YOLO is not about unveiling a novel detector but rather a strategic enhancement using a series of techniques that collectively improve performance, such as replacing the YOLOv3 backbone with ResNet50-vd-dcn, implementing DropBlock regularization, and incorporating IoU awareness. These improvements result in a boost in mean average precision (mAP) on the COCO dataset and increased frames per second (FPS) during inference, outperforming YOLOv4 and EfficientDet. However, while PP-YOLO offers promising results, it is still a new framework, and further empirical testing is recommended to determine its efficacy compared to other detectors like YOLOv5. The ongoing development in object detection, including PP-YOLO's contributions, highlights the potential for further advancements in the field.