Your Computer Vision and Machine Learning R&D lab
Digital Sense is a highly specialized Computer Vision and Machine Learning R&D studio with 12 years of experience and a world-class team of subject matter experts that have published over 175 peer-reviewed papers within a unique research & business perspective.
We offer consulting and development services to companies facing complex challenges in need of solutions with a significant R&D component. We apply our expertise to solve problems in a range of industries, such as AgroTech, Satellite Imaging, FinTech, Digital Health or Beauty, among others.
Our team is helping innovative startups and IT firms to revolutionize their industries within state-of-the art R&D solutions.
CLIENT: A Fortune 500 company, Ulta Beauty is the largest beauty retailer in the United States, operating 1,254 retail stores across all 50 states, with 18,000 full-time employees and an annual revenue of 6.72B in 2019.
CHALLENGE: 8 in 10 women see choosing a cosmetic as a problem while 9 in 10 change products if it does not meet their expectations, but there was no way of trying makeup anywhere, anytime to find the perfect product for customers without them having to actually try it in the traditional way.
SOLUTION: Our team has been working for years with GlamST (a startup acquired by Ulta in 2018) in the development and evolution of their augmented reality makeup try-on solution, which provides a realistic real-time makeup experience to customers. We now continue to work with the Ulta team, our main contribution being R&D incorporated into computer vision and video processing software modules within their platform. Since the onset of the coronavirus pandemic it's become of paramount importance to provide safer experiences to customers. In this context, Ulta Beauty App usage has increased to more than 6 million try ons per month.
CLIENT: Satellogic is a startup with over $60 million of investment that designs, builds and operates a growing constellation of nanosatellites. It has offices in the US, China, Israel, Spain, Argentina and Uruguay.
CHALLENGE: Processing huge volumes of image data obtained with Satellogic´s custom satellite design. This processing needs to be both highly efficient and accurate for clients to quickly receive products they can trust.
SOLUTION: We collaborate with Satellogic's in-house experts, complementing and augmenting their team. We have worked in several stages throughout their proprietary image processing pipeline, from raw images until deliverable products are composed. Main areas of R&D include image geometric and radiometric accuracy, automatic quality assurance and processing efficiency.
CLIENT: Centre National D'Études Spatiales (CNES) is the French National Space Center. It has a budget of €2,780 million and more than 2,350 employees, most of them engineers.
CHALLENGE: CNES is in charge of the development, deployment and operation of several satellites, among them the Pleiades constellation. This is formed by two twin very-high-resolution satellites: 50 cm for panchromatic and 2m for the multispectral bands. Because of their size, images are compressed before sending to the Earth. This compression generates artifacts in the resulting images that degrade the overall image quality
SOLUTION: We performed a profound analysis of the image formation and compression algorithm. From there, we proposed a new mathematical model to implement a joint denoising and decompression algorithm that removes those artifacts. The novelty of the proposed approach led to its publication in one of the best known Image Processing Conferences: IEEE ICIP in 2017. An improvement based on a deep learning approach was later presented at CVPR 2018.
CLIENT: Montes del Plata produces Eucalyptus cellulose pulp. Founded in 2009 by two world-leading forestry companies: Stora Enso (Sweden-Finland, 25,000 employees in more than 30 countries, with EUR 10.1 billion in sales in 2019) and Arauco (Chile).
CHALLENGE: Find a solution for efficient forestry irrigation. The mortality of new trees in drier months is so high as to prohibit planting unless they are watered. The company has watering machines, but they are manually-operated and can only water one row of trees per pass.
SOLUTION: Our team developed a computer vision module mountable on the irrigation machines. It detects plants in real time and automatically activates the mechanism to water trees as the machine advances. This solution enables the implementation of multiple independent watering mechanisms so several rows can be irrigated per pass, saving time, operational costs and improving survival rate over manual operation.
CLIENT: Ministry of Livestock, Agriculture and Fisheries of Uruguay. Uruguay’s agricultural activity represents 25% of the country’s GDP and almost 80% of its exports.
CHALLENGE: Prevent soil erosion and improve the sustainability of the different agricultural activities in the country. This project was executed in the context of a national program.
SOLUTION: Our team built a web Geographic Information System (GIS) with strong satellite imagery processing capabilities as well as mathematical models to compute soil erosion estimations. All of this with the aim of aiding producers in the elaboration of their soil management and crop rotation plans, and to control their execution. The aim of the program is that erosion will be reduced by 80%. Our system is used to oversee all rainfed industrial agriculture in Uruguay: over 3.7 million acres.
CLIENT: Prepaid2cash (P2C) is a startup that for a small fee lets users cash in prepaid and gift cards. Users simply scan their cards and receive in their bank accounts the amounts. CHALLENGE: P2C needed to be able to read both embossed and flat cards (cards with printed digits instead of embossed ones). They had identified an open source library that could be used to read embossed cards but that was failing to read flat ones. SOLUTION: Working with our friends at DecemberLabs, a UX/UI and development company, our team built for P2C a custom component to perform the OCR in flat cards using deep learning models and image processing. This solution is more flexible and accurate, it can read a much broader range of printed cards, and is faster than the pre-packaged version.