February 2021 Summaries

Matt Brems discusses the integration of Roboflow and Streamlit to develop a computer vision app for blood cell count detection, detailing the process from dataset selection to model deployment. Roboflow streamlines the creation of object detection models by offering tools for image uploading, annotating, preprocessing, and augmenting, which simplify the traditionally complex task of training computer vision models even for those without extensive coding experience. The app, created using Streamlit, allows users to visualize predictions of a model trained to detect red blood cells, white blood cells, and platelets, offering functionalities such as adjusting confidence levels and providing immediate visual feedback. Brems emphasizes the ease of deploying such applications with Streamlit, which turns Python scripts into interactive apps, and highlights the importance of evaluating model performance using metrics like mean average precision. This approach aims to empower users to make informed decisions quickly by leveraging the strengths of both platforms in tackling computer vision challenges.

Google Earth Engine, a potent tool for geographic data analysis, is utilized by machine learning experts to examine physical attributes and patterns in various fields such as forestry and weather. Despite its capabilities, there is limited documentation on using its Python API. The process to access and use Google Earth Engine involves authenticating with Google Earth and Google Drive, accessing the built-in IDE or Python API, and understanding the structure of satellite and aerial photographs that compose Google Earth. The guide provides a step-by-step methodology to export images of Des Moines, Iowa, from the Sentinel-2 satellite, focusing on selecting the least cloudy images from 2020 and exporting specific regions to Google Drive. By following the provided notebook and instructions, users can generate and export satellite images ready for use in Roboflow, a platform for building computer vision models.

Mehek Gosalia, a high school student from Sammamish, Washington, developed a disability-accessible rhythm education tool named the Mehek Box, initially as a physical device using tactile blocks to represent rhythmic notes, which she later expanded to include Braille for accessibility. During the pandemic, she further innovated by creating app and web versions that utilize audio, vibration, and animation to provide a multisensory learning experience. Gosalia is collaborating with various programs to test and refine these tools, particularly focusing on bridging the physical and digital versions of the Mehek Box. She introduced an image recognition feature in the app, using a Tensorflow object detection model trained on a custom dataset of rhythm blocks, subsequently augmented with Roboflow's tools to expand the dataset to 10,000 images. Testing different datasets showed surprising results where the unaugmented dataset performed best, leading to further plans for enhancement, such as adding an alignment frame onscreen to assist users in capturing images accurately.

Liquid neural networks, developed by a team of researchers at MIT, represent a significant advancement in artificial intelligence, particularly in the realm of recurrent neural networks. Unlike traditional deep learning models that require retraining and redeployment for new scenarios, liquid neural networks continuously learn and adapt as new data is introduced, making them particularly effective for time series applications such as human gestures, traffic, and power consumption. This flexibility could eventually allow for more resilient models in video processing that adapt to changing environments, although their application in vision tasks is still in the preliminary stages. The promise of these networks lies in their potential to reduce the need for constant active learning and retraining, but it will take further research to extend these capabilities to image and video processing, as current implementations primarily focus on time series data.

During Super Bowl LV, Mountain Dew aired a commercial featuring John Cena in a Mountain Dew-themed amusement park, challenging viewers to count the exact number of bottles in the ad for a chance to win $1 million. To tackle this task, the blog describes using computer vision technology to train an object detection model capable of identifying bottles throughout the scene. The process involved creating a dataset by extracting frames from the video, annotating the bottles in each frame, and then employing image augmentation to enhance the dataset's volume and variability. The model was trained using transfer learning starting from the COCO Dataset, resulting in a model with 89.5% mean average precision. The trained model was then used to analyze the original video, providing a method to accurately count the bottles and potentially win the contest.

James Nitsch, a mobile developer, explores how computer vision can address the challenges associated with grading collectible cards, a process that is traditionally expensive and time-consuming. By developing a proof-of-concept model using images of Pokémon cards, Nitsch demonstrates that computer vision techniques can assist in evaluating card conditions by detecting features such as edge wear, corner wear, and damage. Leveraging tools like web scrapers built in Python and Roboflow's image annotation and augmentation utilities, the model aims to streamline the grading process, potentially offering amateur collectors a more accessible way to assess card value and authenticity. Despite the model's ongoing development, the goal is to deploy it via an API to allow users to upload card images and receive quality and pricing estimates. This approach could significantly reduce the reliance on slow and costly professional grading services, enabling collectors to better navigate market trends and valuation surges.

This blog post provides a comprehensive guide on training and deploying a custom license plate detection model to Luxonis OAK devices using Roboflow and DepthAI. It outlines the process of gathering a dataset of annotated license plate images, leveraging Roboflow for data processing and augmentation, and using Google Colab with TensorFlow for model training. The tutorial emphasizes the flexibility of the approach, allowing users to adapt the steps for detecting different objects. After training, the model is exported and converted into formats compatible with DepthAI and OpenVino before being deployed to an OAK device. The post concludes by encouraging users to enhance their models through additional data collection and retraining, highlighting active learning as a key strategy for improving model performance in production environments.

In February 2021, Roboflow introduced several new features and enhancements across its platform, focusing on areas such as collaboration, data management, and model training. Key updates included the launch of transfer learning and Label Assist, which allows model-assisted labeling using Roboflow Train models. The company expanded team collaboration capabilities by allowing up to three people on the Starter Plan to work together for free and promoted the Upload API to general availability. Other enhancements included support for new data import formats, an object count histogram for dataset health, and improvements to the training process, such as integrating COCO weights for transfer learning and enabling automatic credit reloads for Pro subscriptions. Roboflow also introduced new demo features and reduced the download size of weights files. Additionally, the company announced its seed funding round, launched a new homepage, added OpenAI's CLIP to its model library, and released new public datasets and annotation formats. The platform's updates were complemented by increased media coverage and a variety of user projects leveraging its technology.