We perform reverse image search, allowing users to utilize uploaded images or their links as search queries instead of keywords. Then we return visually similar images (in terms of color, pattern and style) by analyzing the pixels inside the image instead of the metadata associated with it.
For even better results, we add automatic object recognition of objects within an image query, background noise removal, as well as human body detection and skin removal for the fashion industry.
When users watch a video, we can surface visually similar products to those found in the video.
Alternatively, we enable advertisers to place relevant contextual ads on videos. We enable that by matching the visual content found in the ad with the visual content found in the video.
By feeding adequate training data, our machines are able to identify, tag and describe the objects found in images, no matter how large the database is.
We are able to identify objects and their attributes in video content, provided we have the video links or files and a clear categorization purpose and training data.
Our infrastructure architecture enables us to automatically scale up to support the indexing and processing of billions of images, without sacrificing on other performance parameters.
Our results are generated in 100 to 200 milliseconds for search based on an existing database item, and up to 1 second for search based on a newly uploaded image.
Our ability to search and find matching and similar results was evaluated at above 90% satisfaction rate by QA analysts from our customers side.
Our service delivery architecture is designed by putting redundancy and failover in place to achieve high availability and ensure good service levels.
As pioneers, we implemented deep learning for visual search in the retail sector since 2012, and expanded its use to new applications over time.
We use the latest type of artificial neural network, inspired by biological processes.
Our machines are self-taught on how to recognize objects and attributes based on examples provided by our customers.
Our algorithms are constantly being improved while performing tasks on large databases.
Our R&D scientists and infrastructure architects work together to achieve the best performance for our B2B customers.
We use multiple queues and parallel processing to ensure the same speed for all concurrent users.
We auto-scale whenever required by using a distributed architecture and cloud-based servers.
We use specialized NVIDIA GPUs for our deep learning algorithms and high performance power.
NExT is the leading research center in the area of multimedia analysis and search established by National University of Singapore, ranked 22nd in the world, and Tsinghua University, ranked 47th in the world.
Our R&D, Infrastructure Architecture and Dashboard teams count 20 talented computer engineers, out of which 10 are PhD scientists.
Guangda Li, our CTO, has been selected by MIT Technology Review - EmTech as finalist in Top Innovators Under 35.
Our scientists have published their research papers in deep learning and computer vision journals.